By Chris Bell, Product Manager
I build product at Light, which means I spend a large part of my week watching finance teams evaluate systems that all describe themselves the same way. Search "AI ERP" and every vendor on the first page has one. NetSuite ships Text Enhance and Bill Capture. SAP has Joule and a family of finance agents. A wave of newer platforms calls itself AI-native. The label now covers systems whose AI drafts email and systems whose AI closes the books, which makes it nearly useless as a buying signal.
So here is the piece I wish buyers had before those calls: what AI in an ERP actually does, why the architecture underneath matters more than the feature list, where the real risks are, and the questions I would put to any vendor, including us.
What an AI ERP is
An AI ERP is a system of record where artificial intelligence performs or accelerates the work the system exists to manage: recording, coding, matching, reconciling, consolidating, forecasting, reporting. Every useful evaluation I have seen turns on one question. Does the AI inform the people who do the work, or does it do the work?
That question splits the market into three generations, all currently sold under one name.
Rules and robots. OCR that reads invoices, rules that auto-code recurring bills, RPA that replays human clicks. This generation is twenty years old, and most "automation" running in finance departments today is still it. It executes what was predefined and halts at the first surprise, so the exceptions became the job.
Copilots. Conversational AI layered on an existing suite. Ask why travel spend spiked and you get a good answer. Reports get summarized, anomalies get flagged, journal descriptions get drafted. I want to be fair to this generation: it is useful. But copilots compress the time it takes to find and describe things, not the time it takes to fix them, and fixing is where the month goes.
Agents. AI that owns a workflow end to end. Give an agent "reconcile this account" and it works the problem the way a staff accountant would: pulls the bank feed, compares against the subledger, finds the gap, chases the missing receipt in Slack, drafts the correcting entry, and posts it if policy allows. This is agentic accounting, and in my experience it is the only version of AI ERP where the economics visibly move.
The work, concretely
I distrust category language, so here is the actual task list where AI earns its keep: reading and coding invoices, receipts and contracts against your history rather than static rules; matching payments to bills and reconciling bank feeds continuously; proposing accruals from vendor history; drafting intercompany eliminations as transactions post; consolidating entities and currencies continuously instead of in the first week of the month; generating contract-based invoices and recognizing deferred revenue; checking expense compliance at submission and chasing receipts in Slack or Teams; explaining flux at transaction level; answering questions against the live ledger; and flagging duplicates, anomalies and unusual vendor behavior before payment.
Every item on that list exists somewhere as a copilot feature that suggests, and somewhere else as an agent that completes. The product name will not tell you which one you are buying. The architecture will.
Architecture decides the ceiling
Two structural facts determine what any AI inside an ERP can ever do, and neither can be fixed with a better model.
Where the data lives. A batch ERP processes on a schedule, so its AI reasons about the business as of the last run. An AI on a real-time ledger reasons about now. For analysis, stale state is an inconvenience. For action it is disqualifying, because no responsible system posts entries against yesterday.
Where the workflow lives. Agents can only own what they can reach. If the ledger is one product, spend another, AP execution a third and procurement a fourth, every workflow crosses a system boundary, and at each boundary the agent stops and a human resumes. This is why I am skeptical of agents retrofitted onto suites assembled by acquisition, and equally of AI-native ledgers that keep their sub-ledgers in partner tools. For an agent to run procure-to-pay or record-to-report end to end, AP, AR, spend, cards, expenses, procurement, consolidation and reporting have to be native to the same ledger it operates. That conviction is most of the reason Light is built the way it is.
What the AI is allowed to touch. In the age of agents, one of the core buying criteria is simply how many tools the ERP gives its AI access to. Count the verbs, not the features. Can it post a ledger entry? Freeze a card? Chase a receipt? Request approval for an intercompany transfer, and on approval execute both sides of it, the elimination entries and the actual cash movement? Every verb the platform does not expose is a place where the agent stops and a human logs into something. A system whose AI can only read is a copilot, whatever the marketing says; the size of the tool surface is the honest measure of how agentic an ERP really is.
There is a fourth property I would add, because I work on it: whether the system learns your configuration. Traditional ERP setup is frozen at implementation and changed by consultants. Astra, our always-on analyst, watches how a team actually works and suggests workflow and configuration changes to increase performance, applied in-product. A system should get better with use, not drift out of date.
The controls question
When I show agents to controllers, the first question is always the same: if agents post entries, what happens to my controls? It is the right question, and I would judge any vendor by the precision of their answer.
The correct design is that agents live inside the same framework as employees. The finance team sets approval thresholds, segregation of duties and posting permissions; work below the threshold flows through; everything else is prepared, documented and queued for approval. Every agent action is logged with what was done, on what evidence, under which policy, in identical structure every time. The audit trail stops being a reconstruction from emails and spreadsheets and becomes a byproduct of the work.
Do not accept this as a slide. Ask to see an individual agent action: attributable, evidence-backed, policy-constrained by amount, account and entity, and reversible.
Numbers I am willing to defend
A lot of percentages in this category are modeled assumptions dressed as results, so I will only cite measured ones from our own customers. Tillo processes $2B in revenue across 8 entities on Light and cut month-end processing time by 76%. Ocean.io replaced QuickBooks, e-conomic and a patchwork of point solutions, cut its close by 60%, and added a second entity without adding headcount. Lovable scaled to $500M in revenue with a finance team of four. Alva Labs went from multiple systems and weekend closes to a near-zero-day close. Biofire moved off NetSuite and runs four US entities on Light with real-time event streaming into its manufacturing systems.
The pattern repeats: the mechanical layer moves to agents, the close compresses from an event into a property of the system, and the team's time moves to judgment.
The questions I would ask any vendor
- Can the AI post an entry, or only suggest one? Insist on a live demonstration, not a roadmap slide.
- How many tools can the AI actually call? Ask for the list of actions, not features: post journal entries, freeze a card, request approval for an intercompany transfer and then execute both the accounting and the cash transfer. The length of that list is the product.
- Where do agents stop? Trace one workflow, invoice to posted payment, and count the system boundaries and human handoffs.
- What state does the AI see at 3pm on a Tuesday? Real time, or last night's batch?
- Show me the audit trail of one agent action. Attributable, evidence-backed, reversible, or a black box?
- Does the system learn my configuration, or does change mean re-engaging consultants?
- Is it global? Multi-entity, multi-currency, local rails and tax logic in the countries I operate, native or via partners?
- What does implementation look like? Weeks with a dedicated team, or a program measured in fiscal years?
Where I would push back on my own industry
Three things deserve more honesty than vendor pages give them. Data quality still gates everything; an AI ERP inherits the mess it is migrated onto, which makes implementation quality matter more, not less. Change management is real work; a team that has spent years as exception-handlers needs help becoming policy-designers. And for very large enterprises with deeply customized processes, the incumbent suites remain a rational choice today. The calculus flips fastest for multi-entity companies from tens of millions to low billions in revenue, where the gap between a week-long close and a continuous one compounds every month. If you want the landscape mapped vendor by vendor, incumbents included, we keep it at AI alternatives to ERP, compared.
Where this goes
Every roadmap in the industry is promoting copilots into agents. The difference is that on an AI-native ledger that promotion is a feature release, and on a batch suite it is an architecture migration. Within a few years I expect the close to disappear as an event at companies running agentic systems, configuration to become something the system proposes rather than something consultants bill for, and "AI ERP" to fade into the assumed baseline, the way "cloud ERP" did a decade ago.
Frequently asked questions
What is an AI ERP? A system of record where AI performs or accelerates the core work: coding, matching, reconciling, consolidating, forecasting and reporting. The distinction that matters is whether the AI suggests work or completes it inside your controls.
What is the difference between AI ERP and AI-native ERP? AI ERP covers any ERP with AI features, including copilots retrofitted onto batch architectures. AI-native means the system was designed around AI doing the work: a real-time ledger, native sub-ledgers so agents can reach the whole workflow, and controls built for attributable agent actions.
Can AI replace an ERP? No. The system of record remains. What changes is the operating model: from humans doing the work and software recording it, to agents doing the work and humans governing it.
Do AI agents break audit and compliance? Not when the architecture is right. Agents operate inside approval thresholds and segregation of duties set by the finance team, and every action is logged with its evidence. The trail gets more consistent, not less.
What results do companies actually see? Measured examples from Light customers: a 76% cut in month-end processing at $2B-revenue Tillo, a 60% close reduction at Ocean.io, and a four-person team scaling Lovable to $500M in revenue.
Chris Bell is a Product Manager at Light, the agentic accounting platform. See the product or book a demo.