On 1 April 2026, Annexa and Oracle NetSuite ran a live webinar covering how AI is showing up inside real finance workflows — from embedded NetSuite capabilities through to agentic AI tools and what’s coming next with NetSuite Next. The session generated a strong volume of questions, many of which we didn’t have time to get through on the day. This post works through all of them.
Watch the full webinar on demand here.
Connecting AI assistants to NetSuite – when will the connector support Copilot?
This question came up early in the session and it’s a good one, because the answer depends slightly on what you mean by ‘the connector’ and which version of Copilot you’re asking about.
NetSuite’s AI Connector Service is built around the Model Context Protocol (MCP) – an open integration standard that allows AI assistants to connect to external systems in a governed, permissioned way. The design intent is explicitly ‘bring your own assistant’, meaning it’s not locked to a single AI provider. Any MCP-capable client can connect to it, provided the authentication and protocol requirements are met.
Microsoft Copilot Studio now supports MCP as a generally available feature – this is no longer in preview. You can connect a Copilot Studio agent to NetSuite’s MCP endpoint using the built-in MCP onboarding wizard, which walks through server URL, authentication and tool configuration. The practical question is which Copilot product and integration path you’re working with – a Copilot Studio agent, a Microsoft 365 Copilot connector or something else – and whether your team is building that connection or waiting for a pre-packaged experience.
One thing worth knowing if you’re connecting ChatGPT rather than Copilot: as of March 2026, ChatGPT changed its Callback URL from static to dynamic, which means existing NetSuite integration records built for ChatGPT need to be recreated. Connections established before March 4, 2026 continue to work, but new connections require a fresh integration record each time.
The short answer is – MCP support on both sides means the connection is achievable. The timeline for a fully packaged, low-configuration Copilot experience inside NetSuite is less certain and worth monitoring through NetSuite’s release notes.
In the above demo, we show how the NetSuite AI Connector allows external AI platforms – including Claude and ChatGPT – to work directly with your live NetSuite data, inside governed, permission-controlled workflows. This demo was recorded as part of Annexa and Oracle NetSuite’s webinar: AI for finance leaders – what to use now and what’s next, held 1 April 2026.
Permissions – what should the AI role in NetSuite actually have?
Before getting into the specifics, it’s worth a quick note on how NetSuite’s access model works. NetSuite controls what every user – human or system – can see and do through a role-based permissions framework. Each role is a collection of permissions that determines which records can be accessed, what actions can be taken and which features are visible. When you connect an AI assistant to NetSuite via the AI Connector Service, it operates under exactly the same framework – it gets a role and that role defines its boundaries.
NetSuite is explicit about what that role needs and the guidance is worth following closely.
The minimum requirements to enable MCP access are: Server SuiteScript and OAuth 2.0 must both be enabled on the account and the role used for AI access needs permission for MCP Server Connection and Log in using OAuth 2.0 Access Tokens. NetSuite specifically warns not to confuse this with ‘Log in using Access Tokens’ – they’re different settings. If you want the AI to be able to create or update records (not just read and query), you’ll also need REST Web Services enabled and permissioned for that role.
Two non-negotiables come through clearly in NetSuite’s own documentation. First, MCP access is off by default – it must be explicitly granted to a role. Second, the AI Connector Service cannot be used with Administrator-level roles. NetSuite blocks this explicitly, which is the right call from a governance perspective.
The cleanest setup is a dedicated AI role that has MCP Server Connection plus only the specific record-level permissions needed for the tasks you intend to expose. Least privilege applies here just as it does anywhere else in your integration architecture – the AI should be able to do exactly what you need it to do and nothing more.
Data security – how safe is your data when using Claude or other third-party AI?
This is one of the most important questions any finance team should be asking before connecting live ERP data to an external AI platform and the answer has two parts: what NetSuite controls and what happens once data leaves NetSuite.
On the NetSuite side, the controls are meaningful. As covered in the AI Connector setup documentation, MCP tools run with the same permissions as the NetSuite user operating the AI agent – so if a user can’t see a record, the AI can’t either. All MCP tool usage is logged, giving you traceability for actions performed by AI agents. There are also hard limits on what MCP tools can do – they can’t run as a different role, can’t invoke scripts with elevated privileges, can’t call Suitelets and can’t make outbound HTTP requests to external destinations. NetSuite also requires explicit user acknowledgement before connecting to a third-party LLM and makes clear that users are responsible for controlling what data is shared.
On the Claude side, Anthropic draws a sharp line between consumer and commercial products – and the distinction matters more than most teams realise. In August 2025, Anthropic updated its consumer terms so that Free, Pro and Max account users are now opted in to model training by default unless they actively turn it off. Data from those accounts can be retained for up to five years if training is enabled. This applies even to paid consumer plans – Claude Pro is a consumer account, not a commercial one.
Commercial products operate under entirely different rules. For Claude for Work, the Anthropic API, Enterprise and similar tiers, Anthropic does not use inputs or outputs for model training by default – no opt-out required. API log data is retained for just 7 days by default. For organisations with stricter requirements, Anthropic also offers a Zero Data Retention agreement for eligible enterprise API customers, under which inputs and outputs are not stored beyond what’s needed to screen for misuse.
Keep in mind that if an employee is using a personal Claude Pro or free ChatGPT account to analyse sensitive ERP data – what’s sometimes called shadow AI – that data may be feeding model training without anyone in the organisation being aware. NetSuite’s own guidance on data security reinforces the importance of controlling what leaves your system. The right safeguard is to establish a clear organisational policy that defines which AI tools are approved for use with finance data and ensures those tools are accessed via enterprise or API pathways – not personal accounts.
It’s also worth noting that these policies change – sometimes with limited notice, as the August 2025 Anthropic update demonstrated. Building a periodic review of your approved AI tools and their current data terms into your governance calendar is a simple step that keeps your organisation ahead of the risk.
Getting the right data – how do you point AI to the right numbers in NetSuite?
Two related questions came through on this – one about how to direct AI to the right data given NetSuite’s complexity and one specifically about a revenue chart that came back with incorrect numbers.
The core issue here is usually not the AI’s capability – it’s definition. When you ask an AI to ‘show me revenue’, it needs to know which subsidiary, which currency, whether you mean recognised or billed revenue, what period and what eliminations apply. If those aren’t specified, the AI makes assumptions and assumptions produce wrong numbers. This is a data and context problem, not a technology one – and it’s worth keeping that framing in mind before assuming the AI is broken.
The most reliable pattern is to separate the data retrieval step from the output step. Ask the AI to pull and confirm the underlying dataset first – using a saved search, a standard report, or a SuiteQL query – and agree on the numbers before asking it to produce a chart or commentary. When the AI is working from a fixed, named artefact rather than interpreting a freeform question, the risk of wrong numbers drops significantly. NetSuite’s MCP Standard Tools support all three retrieval methods, so you can direct the AI to something specific and verifiable rather than leaving it to guess.
NetSuite has also recently addressed this problem directly with the AI Connector Service Companion, announced at SuiteConnect London in March 2026. It includes a library of over 100 finance-specific prompt templates aligned to NetSuite’s data structures, permissions and terminology, organised by business process and recommended role – covering positions like CFO, Controller, AP Analyst and Treasury Analyst. Users can also customise existing templates or add their own. This is a practical response to exactly the problem described above: most wrong-number issues are definition problems and standardised templates reduce the trial-and-error that comes with ad hoc querying. If your team is regularly using AI against NetSuite data, this is worth exploring as a starting point rather than building prompts from scratch.
For a broader overview of what’s now possible with AI and NetSuite data, the Annexa blog covers the 2026.1 release in detail.
Hallucinations – how do you validate that AI analysis is actually correct?
This is the right question to be asking and the fact that it came up multiple times during the session suggests finance teams are approaching AI with appropriate scepticism.
NetSuite’s own risk documentation explicitly names hallucination as a known weakness of large language models – defining it as situations where the AI generates information that appears accurate but is in fact incorrect or fabricated. Importantly, NetSuite states directly that hallucination is outside of its control and advises users to always validate AI outputs against source data. When the vendor building the platform is telling you to check the numbers, that’s the clearest possible signal that validation needs to be part of your process.
The same documentation also flags prompt injection as a related risk – where hidden instructions embedded in content processed by the AI can cause it to perform unintended actions. Both risks are worth understanding before connecting AI to live finance data and both are managed through the same principles – limited permissions, trusted tools and validation of outputs before acting on them.
In practice, the most effective mitigation for hallucination is to use structured retrieval tools – saved searches, reports or SuiteQL queries – to get a verifiable dataset first, then use the AI to interpret and communicate from that agreed foundation. When the AI is working from a fixed, queryable dataset rather than generating numbers from context, the hallucination risk on numerical output drops significantly. The key discipline is not skipping the verification step, even when the AI’s output looks plausible.
NetSuite’s 2026.1 update introduces Ask Oracle as part of NetSuite Next, which takes a related approach to this problem by providing reasoning that explains the ‘how’ and ‘why’ behind responses, grounded in existing roles, permissions and data. The transparency of that reasoning chain is itself a validation mechanism – if the AI can show you what it queried and how it arrived at a number, you can follow the logic and spot where it may have gone wrong.
Integration effort – how much work is it to give AI access to NetSuite?
Less than most teams expect – and the starting cost is zero. The AI Connector Service and the MCP Standard Tools SuiteApp are both free, which removes what many assume is the first barrier.
The setup itself is largely configuration rather than development. You install the MCP Standard Tools SuiteApp from the SuiteApp Marketplace, configure a role with the permissions covered earlier in this post and connect via NetSuite’s account-specific MCP endpoint using OAuth 2.0. You’re not giving an AI direct database access – you’re enabling a governed API layer that the AI calls through, with the same permission controls that govern the rest of your NetSuite integrations. The AI discovers the available tools automatically once connected – there’s no manual configuration of endpoints or data schemas required.
Two practical things are worth knowing before you go live. First, the AI Connector Service draws from the same concurrency pool as your other integrations. If your account has a total concurrency limit of five requests and two are already allocated to another integration, the AI Connector can only use the remaining three. If you exceed that limit, you’ll see a ‘Too Many Requests’ error and need to retry. It’s worth reviewing your current concurrency allocation before connecting AI workflows, particularly if you’re already running other active integrations. Second, execution logs for AI Connector Service activity are retained for 21 days in production and seven days in sandbox – build your monitoring and troubleshooting practices around that window.
For those already on NetSuite, the 2026.1 release blog covers what’s changed in the latest release and what to expect next. If you’d like a hand scoping the setup for your environment, get in touch with the Annexa team directly.
AP automation with Erstan – email inboxes and batch uploading
Two specific questions came through during the AP automation demo – one about whether bills can be submitted via email and one about whether files need to be processed one at a time. Worth clarifying what was shown in the session: the demo below featured Erstan, Annexa’s purpose-built agentic AI platform for NetSuite, handling AP automation as one of its core workflows. This is distinct from NetSuite’s native Bill Capture feature, so it’s worth explaining how Erstan’s AP agent actually works before answering both questions.
Erstan’s AP Vendor Bill Automation agent monitors an email inbox automatically. When a vendor bill arrives, the agent classifies the email, kicks off processing and loops you in via notification when it needs your approval. You can also trigger it manually by uploading a bill via form or chat. Either way, the agent handles extraction, vendor matching, PO lookup, duplicate detection, bill creation and source document attachment – with a single approval gate before anything is written to NetSuite.
On the email inbox question specifically yes, Erstan monitors your inbox directly. The agent processes vendor bills the moment they land, without anyone needing to forward or upload them manually. The original PDF is automatically uploaded and attached to the created bill in the NetSuite file cabinet – so there are no lost invoices and no manual filing step.
On batch processing: because Erstan monitors the inbox continuously rather than waiting for a manual upload, bills don’t queue up waiting to be processed in batches – they’re handled as they arrive. For manual uploads, bills can be submitted individually via form or chat and processed immediately. The agent runs a single-pass lookup covering vendor match, PO match, bank details match and duplicate check before presenting the full findings for your review and approval.
What makes Erstan’s approach different from standard NetSuite AP automation is the agentic layer on top. Where NetSuite’s Bill Capture uses OCR to extract data and populate a draft for review, Erstan uses what it describes as agentic OCR – a goal-oriented process where the system reasons about document structure, chooses the right model for each page section and runs a self-correction loop before passing results downstream. The extracted totals are also verified against the created bill before anything reaches the ledger, with any variance flagged before posting.
For teams managing high volumes of vendor bills or wanting inbox-triggered automation with a full audit trail, get in touch with the Annexa team to find out more about Erstan.
How does Erstan compare to third-party AP automation tools?
This is a fair question to ask, and the honest answer is that it depends on what you’re comparing and what you’re trying to solve.
Most third-party AP automation tools on the market are built to solve a specific problem: getting invoices out of inboxes and into your ERP with less manual effort. They typically do this well for high-volume, standardised invoice flows – particularly where vendors send consistent formats and the matching logic is straightforward. The trade-off is that they sit outside your ERP, which means data moves between systems, audit trails are split across platforms, and any exception handling that requires ERP context involves a hand-off.
Erstan is built differently, and the difference starts with architecture. Because Erstan connects directly to your live NetSuite environment rather than integrating via middleware, every step of the AP process – vendor matching, PO lookup, duplicate detection, bill creation, document attachment – happens inside your NetSuite data model. There’s no export, no sync, no reconciliation between systems. The original PDF is attached to the created bill in the NetSuite file cabinet automatically, linked to the transaction it created.
The second difference is what Erstan means by automation. Traditional AP tools use OCR to extract invoice data and populate fields for review. Erstan uses what it calls agentic OCR – a goal-oriented process where the system reasons about document structure, chooses the right approach for each section of the document, and runs a self-correction loop before passing results downstream. Extracted totals are verified against the created bill before anything reaches the ledger. Any variance is flagged before posting. That reconciliation guarantee is built into the workflow rather than left to a downstream review step.
The third difference is the audit trail. Every Erstan agent run produces a full step-by-step trace – what data went in, what rules were applied, what checks passed, and what was written to the ledger. When an auditor asks how a bill was created, you can show them exactly what happened at every step. Most standalone AP tools don’t produce that level of traceability because they weren’t designed with finance-grade audit requirements in mind.
The fourth difference is scope. Erstan’s AP automation agent is one of 50+ agents available out of the box, covering AR, month-end close, financial reporting, compliance and audit. A standalone AP tool solves one problem. Erstan is a platform – so the investment in connecting and configuring it extends across your finance operation rather than being limited to invoice processing.
Where third-party tools may still make sense is in environments with very high invoice volumes, complex supplier portals, or specific EDI requirements that sit outside NetSuite’s native capabilities. For teams on NetSuite looking for governed, auditable AP automation that lives inside their existing environment rather than alongside it, Erstan is purpose-built for exactly that.
Erstan is currently in beta and open for early access. Register your interest here and the Annexa team will be in touch.
Get a personalised NetSuite pricing estimate
Receive a tailored NetSuite quote based on your business requirements, users, modules and implementation needs.
NetSuite Next – change management, customisations and rollout control
Several questions came through about NetSuite Next – specifically around whether teams can control when it’s turned on, whether it will affect existing scripts and customisations and what preparation is needed. But first, let’s refresh on what NetSuite Next actually is – it’s Oracle’s most significant platform overhaul since NetSuite’s founding, bringing a new Redwood UI, the Ask Oracle natural language assistant, AI Canvas for scenario planning and agentic workflows. Annexa covered the SuiteConnect 2026 announcements in detail here if you want the full picture.
On rollout control, the transition is designed to be gradual and opt-in rather than forced. The Redwood Experience UI – the visual layer of NetSuite Next – is already available and can be enabled per user by going to Home > Set Preferences > Appearance. Oracle has also provided the option to toggle it off at the user level, which means teams can run a pilot group before committing to a broader rollout. Ask Oracle and AI Canvas are rolling out through 2026 – currently North America first, with other regions including ANZ following. Agentic features are in preview or early access as of early 2026, with Oracle releasing them incrementally rather than all at once. For ANZ customers, the practical approach is to enable Redwood now for willing users, monitor Oracle’s release schedule for Ask Oracle availability in your region and plan broader adoption once you’ve tested in your environment.
On customisations and scripts, Oracle’s stated position is that NetSuite Next does not require migration and should not disrupt existing customisations. Ask Oracle is described as operating across customisations and extensions built on SuiteCloud, including partner applications. That said, a UI change of this scale can surface layout and workflow assumptions that weren’t visible before. Any team with meaningful SuiteScript, Suitelets or heavily customised forms should plan regression testing in a release preview or sandbox environment before rolling out broadly – even if Oracle’s messaging is that the switch is non-disruptive.
On costs, NetSuite Next is included in existing licenses at no extra charge. The cost exposure sits elsewhere – in the change management, training, control revalidation and governance work that any significant platform update requires. If your team has complex customisations or a large user base, that investment is worth planning for explicitly rather than treating the rollout as a purely technical exercise.
Which AI tools should you use and which should you avoid?
This question included an example of a tool giving incorrect calculation results and attributing it to a ‘mis-type’. That’s a hallucination and it’s worth naming it as such – AI tools make mistakes, including well-regarded ones and numerical reasoning is one of the areas where errors show up most often. Choosing the right tool matters, but so does understanding what any tool is and isn’t suited for. Annexa doesn’t prescribe specific platforms – the right choice depends on your environment, your existing systems and what your team has already standardised on. The clearest framework for evaluating any AI tool for finance use is to ask four questions.
Is training enabled by default? This is the single biggest risk factor for finance teams. As covered in the data security section above, consumer accounts on platforms like Claude Free, Pro and Max and standard ChatGPT accounts can now be set to train on your conversations unless you actively opt out. Any tool where training is on by default – or where your team may have already accepted terms without realising it – is a governance risk when used with sensitive financial data. Enterprise and commercial accounts on the same platforms operate under entirely different terms, with training off by default. Audit what your team is using before assuming you’re protected.
Does it integrate with your systems in a governed way? Tools that connect directly to your NetSuite environment through the AI Connector Service operate within NetSuite’s existing permission and audit framework – the AI can only see what the user can see and every interaction is logged. Tools that require you to export data out of NetSuite and paste it into a chat window introduce data handling risks that sit entirely outside that framework. The gap between those two approaches is significant.
Is it built for the task at hand? General-purpose AI tools are well suited to drafting commentary, summarising reports and answering ad hoc questions. For structured finance workflows – journal creation, reconciliation, AP processing, forecasting – they introduce the kind of unstructured, unpredictable outputs that finance controls are designed to prevent. The tools that work best in those environments are purpose-built for them, with defined workflows, validation steps and audit trails rather than open-ended prompting.
Can you verify the output? Whatever tool you use, the ability to trace how an output was produced matters. If an AI gives you a number and you can’t see what it queried or how it calculated the result, you’re relying on trust rather than verification. Prioritise tools that show their reasoning and link outputs back to source data.
For finance teams on NetSuite looking to move beyond ad hoc AI use and into structured, governed workflows, Erstan is Annexa’s purpose-built platform for exactly that – agents that execute real finance processes end to end, with human approval gates, full audit trails and direct links back to NetSuite records. At the time of writing Erstan is in beta and open for early access. You can register your interest for launch here.
Will NetSuite users start bypassing off-the-shelf FP&A tools?
This question gets at something a lot of finance leaders are genuinely thinking about as AI capabilities inside ERPs improve.
The honest answer is: sometimes, for some use cases. If your planning needs are relatively straightforward and your main pain is getting faster answers from your NetSuite data, a well-configured AI Connector with good prompt templates can genuinely reduce the case for a standalone FP&A tool. The gap between ‘export to Excel, analyse, bring back’ and ‘query NetSuite directly via AI’ is narrowing – and NetSuite’s 2026.1 release accelerates that further with AI payment date prediction, enhanced bank reconciliation matching, and the new Intelligent Close Manager.
Where dedicated FP&A tools remain relevant is in more complex planning environments – driver-based modelling, governed forecast versions, cross-entity consolidations, audit trails and repeatable planning cycles. NetSuite’s own answer to this sits in its EPM stack. NetSuite Planning and Budgeting handles budgeting, forecasting and scenario modelling with embedded AI and ML for trend analysis, forecast refinement and AI-generated commentary. NetSuite Enterprise Performance Management goes further – adding account reconciliation, financial consolidation, close management and narrative reporting, built on Oracle Fusion Cloud EPM with native NetSuite integration. The 2026.1 release added AI agents across all three EPM modules – Account Reconciliation, Planning and Budgeting, and Profitability and Cost Management – making the case for the EPM stack stronger than it was even twelve months ago.
It’s worth flagging that NetSuite EPM is a separately licensed product. It’s not included in a standard NetSuite subscription, so if EPM capabilities are relevant to your planning needs, that’s a conversation worth having with your Annexa account manager before assuming it’s already available in your environment.
The strategic question for most teams isn’t ‘ERP AI vs FP&A tool’ – it’s which planning complexity level you’re operating at, and whether your current tooling is genuinely serving that need or just familiar. If you’re not sure where you sit, the Annexa team is happy to walk through the options.
Can AI support a NetSuite implementation?
Yes – and it’s already happening across several parts of the implementation process, though it’s worth being clear about where it adds genuine value versus where human expertise is still doing the heavy lifting.
The areas where AI is making a real difference in implementations right now tend to be the more time-intensive, document-heavy parts of the process. Requirements gathering and documentation – summarising workshop outputs, drafting process specifications, and turning meeting notes into structured requirements – moves faster with AI assistance. Data migration preparation, including reviewing legacy data for gaps, inconsistencies and mapping issues, is another area where AI can cover a lot of ground quickly. The same applies to test script generation, training material drafting, and producing first-pass documentation for custom configurations.
For teams already on NetSuite looking to optimise their environment, AI tools connected via the AI Connector can also support configuration reviews – querying live data to surface unused workflows, dormant roles, or process gaps that might otherwise take days of manual analysis to identify.
What AI doesn’t replace is the judgement that comes with implementation experience – understanding how a business actually operates, what will and won’t work in practice, and how to design a NetSuite environment that supports growth rather than just replicating what was there before. That’s where Annexa’s implementation team brings the most value, and where the combination of experienced consultants and AI-assisted tooling is producing better outcomes than either alone.
If you’re planning a NetSuite implementation or looking to get more from an existing environment, get in touch with the Annexa team to talk through what’s involved.
Is there a roadmap for AI across NetSuite Analytics Warehouse, FSM and EPM?
Yes for NSAW and EPM – and it’s moving quickly. FSM is a different story, so it’s worth addressing each separately.
NetSuite Analytics Warehouse has seen some of the most significant AI additions in the 2026.1 release. The headline addition is the AI Connector Service for NSAW – which extends MCP-based AI access beyond live transactional ERP data to the historical, analytical and third-party data stored in the warehouse. That’s a meaningful extension of what you can query via AI, opening up longer-horizon analysis, cross-system datasets and richer forecasting inputs that weren’t previously accessible through the connector. The release also introduced the Oracle AI Database 26ai with performance enhancements, and NSAW already includes prebuilt AI and ML models for pattern discovery, anomaly detection and predictive analysis – with more models in development for use cases like customer churn and inventory stockouts.
NetSuite EPM gained AI agents across three modules in 2026.1 – Account Reconciliation, Planning and Budgeting, and Profitability and Cost Management. The reconciliation additions are particularly significant: a new AI assistant automates reconciliation setup by identifying newly created accounts, assigning preparers, applying formats and learning from prior cycles, while an AI-powered transaction matching assistant supplements existing rule-based engines with machine learning that catches patterns traditional rules miss. On the planning side, embedded AI and ML continuously analyse data trends and drivers, and AI-generated commentary explains why numbers change rather than just reporting that they have. If you’re on NetSuite EPM, the 2026.1 release makes a strong case for reviewing which of these capabilities are active in your environment – and if you’re not yet on EPM, it’s a reason to evaluate it.
FSM is the exception here. The 2026.1 release notes for Field Service Management cover mobile sync improvements and interface localisation – useful operational updates, but no meaningful AI additions. If FSM is a priority area for your organisation, it’s worth raising with your Annexa account manager so they can keep you informed as the roadmap develops.
The broader direction across the platform – as covered in the Annexa SuiteConnect 2026 recap – is AI embedded directly into finance workflows on one track, and governed external AI connections via MCP on the other. Both are active and accelerating, with the pace of announcements in early 2026 suggesting this is a sustained investment rather than a short-term push.
Want to continue the conversation?
If any of these questions connect to something you’re working through in your own environment, the Annexa team is happy to walk through the specifics. Whether you’re on NetSuite already or evaluating it or thinking about where to start with AI in your finance workflows, get in touch.
Contact us at annexa.com.au or call 1800 319 685.