That includes the general ledger and the newest applications, PSA and Inventory.
In an earlier piece we explained in plain words what an API is and why AI makes it matter. This article is the concrete follow-up. I want to show what our APIs actually cover, how they’re built, what an AI agent can really do with them, and where we differ from newer finance platforms such as Rillet and Campfire.
The short version: for twenty years the question was “does your ERP have an API?” Today the question is “can an AI agent run a real business process in your ERP, under the same rules as your accountants?” Those are very different questions.
Key takeaways
Lensing’s API catalog covers about 580 actions and 180 data sets across eleven application areas, all available over REST and through each customer’s own MCP server.
AI agents use the same API contract as Lensing’s own interface, so they follow the same permissions, approval policies and accounting rules as people.
Agents can complete real workflows end to end, from month-end accruals to AR aging and inventory landed costs, while people keep the approvals.
You can [try it yourself in a free sandbox](LINK: MCP page) with ChatGPT or Claude.
Why APIs matter more in an ERP than anywhere else
An ERP is not just a database with screens. It is the system of record, and the API is the place where the record is protected.
Three things make ERP APIs different from the APIs of most business tools:
Every write has accounting consequences. A vendor bill touches an entity, a book, a period, a currency, a tax code, an approval policy and, eventually, the general ledger. An API that just stores rows lets anyone break the books. An API that enforces the rules keeps them intact.
Coverage is the job description of an AI agent. An agent can only do what the API lets it do. If revenue recognition, inventory or project billing are missing from the API, the agent stops there and a person takes over with exports and spreadsheets.
Agents need more than endpoints. A developer reads documentation once and remembers it. An agent discovers the system at runtime. It needs precise descriptions, typed schemas and error codes it can act on. That is what MCP adds on top of the API: the system describes itself to the model.
So the right questions are: how much of the ERP does the API cover, how well does it enforce the rules, and how well does it explain itself to a machine?
What “complete APIs” means at Lensing
Our API catalog today holds about 580 actions and 180 queryable data sets across eleven application areas. Every one of them is reachable over REST and over the [MCP server](LINK: MCP page) of each customer tenant.
Actions are the verbs: create, update, approve, commit, post, reverse, record. Data sets are the nouns you can query, filter and page through, down to the line level.
Two details matter more than the count. First, the APIs reach the line grain: invoice lines, journal entry lines, revenue plan lines, payment applications. That is where reconciliations and analyses actually happen. Second, they cover the full lifecycle of a document, not just its creation. A time entry can be submitted, approved and rejected. A warehouse transfer can be committed and billed intercompany.
One contract for people, integrations and AI agents
In Lensing, the API is not a side door next to the application. Our own React UI talks to the platform through the same OData APIs. Whatever rule applies to an accountant clicking a button applies to an integration and to an AI agent in exactly the same way.
Take the simplest finance action there is: creating a manual journal entry. Through the API, this is what happens:
The entry is created with its lines, unposted and pending approval.
If an approval policy applies, it routes to the approvers. Only approval posts it to the general ledger.
If an accounting rule rejects it, for example because debits and credits don’t balance, it is not created at all. The caller gets a VALIDATION_FAILED error with the rule that failed.
Every line must reference an existing account and legal entity. Foreign-currency lines take the latest exchange rate on or before the posting date unless a rate is given.
Lines can carry reporting tags across nine system dimensions (department, business unit, team, location, region, employee, vendor, customer, value center) and any custom dimension. A tag pointing to an inactive record is refused.
With multi-book accounting, the entry can be derived into every book in one call.
None of this is special code for AI. It is the same contract the UI uses. That is the point: an agent cannot take a shortcut that a person could not take.
The errors matter as much as the successes. Each action declares its error codes and their payload. An agent that gets REFERENCED_RECORD_NOT_FOUND knows it has to look up the department first. It does not have to guess from a generic HTTP 400.
What can an AI agent do in an ERP end to end?
Coverage only matters if it lets an agent finish a real process. These are workflows that run against Lensing’s APIs today, from a general-purpose assistant such as Claude connected to the tenant’s MCP server.
The pattern is the same in every row. The agent does the reading, matching and drafting at machine speed. The approval policies the company configured still decide what reaches the books.
It also works the other way round. Because the APIs are complete, a customer can build their own application on top of Lensing with an AI coding assistant, for example a cash-flow view or a treasury tool after a merger, without waiting for us to ship it. One of our own new hires, with no ERP experience, [built a travel and expense app this way](LINK: Gereon’s article).
Built for AI agents at scale
People make mistakes too. The difference is that a thousand agents make them a thousand times faster. So the platform under the API has to be designed for many concurrent, automated callers, not for a handful of users in a browser.
These are the properties we built in for that:
Permissions per API action. Every action and data set is a permission target. An agent acts with the rights of the user or role it runs as, nothing more. If access is missing, the API says which role or task would grant it.
Transactions that hold under concurrency. Conflicts are detected at database level and retried automatically where that is safe. Explicit named locks protect the cases where it is not.
Nothing is overwritten. Persistence is insert-only with version intervals. Every change an agent makes stays visible, with who made it and when.
Events instead of polling. Consumers register for changes and receive them through a transactional outbox, delivered at least once. An agent can react to a new vendor bill instead of asking every minute.
Sandboxes for trying first. [Live Sandboxes](LINK: Live Sandbox page) let an agent or a developer work on a copy of the tenant. The application log is replayed when the sandbox is merged.
Stable, versioned contracts. Customer extensions and agents build on fully versioned APIs. Our bi-weekly releases do not break them.
Isolation and data residency. Each customer has its own tenant and its own MCP server, operated in the US or in Europe. AI can be restricted to models that keep data in the region and do not train on it.
How Lensing’s MCP compares to Rillet and Campfire
Rillet and Campfire are serious products, and both deserve credit for shipping MCP early. Both are built as AI-native general ledgers for high-growth SaaS companies. Lensing is built as a complete ERP. That difference shows up directly in what an AI agent can reach.
All three also run their own in-product AI agents that can connect outward to other tools. This comparison is about the opposite direction: what your own AI assistant, such as ChatGPT or Claude, can do inside the ERP through its MCP server.
Three differences follow from this.
Breadth. Rillet’s MCP is strong on SaaS billing, revenue and core AP/AR, which is its core. Campfire’s is a reporting and analysis interface. Neither covers inventory, project-based services, purchase orders or HR, because these are not their products. If your company sells hardware with software, runs services projects or manages a warehouse, the agent stops at the edge of the ledger.
Depth of control. A read-only MCP is a very good analyst, but it cannot do the work. A write-capable MCP is only as safe as the rules underneath it. In Lensing, the MCP server uses the same API contract as our own interface: every write goes through the user’s permissions, the company’s approval policies and the accounting rules, and nothing is overwritten. Every change stays visible, with who made it and when. An agent cannot take a shortcut a person couldn’t.
Globalization. Lensing is built to meet each country’s legislation and expectations, with GoBD-compliant document storage, e-invoicing and local charts of accounts, operated in the US or in Europe. An agent posting a German vendor bill works with the same rules a German accountant would.
This comparison is based on each vendor’s published MCP documentation as of November 19, 2026. All three products move fast, and their scope will grow.
Five questions to ask any ERP vendor about AI and APIs
If you are choosing a finance system in the age of AI, “do you have an API?” is no longer a useful question. These are:
Which share of your applications can an agent read and write through the API, and which processes end at the edge of it?
Do API calls go through the same approval policies and accounting rules as the UI, or around them?
Does the API explain itself to a machine, with typed schemas, descriptions and specific error codes?
Can I see every change an agent made, and try things in a sandbox first?
Can I build my own applications and agents on it without breaking at the next release?
We built Lensing so that the answer to all five is yes. The API is no longer a feature for developers. It has become the most important user interface of the ERP.
See it for yourself. [Get a free Lensing sandbox](LINK: MCP page), connect ChatGPT or Claude, and put an agent to work on a real finance process. Prefer a guided tour? [Join our live product demo](LINK: weekly demo), every Thursday at 10am PT.
Sources
Lensing API catalog, tenant catalog query, November 2026
Rillet: How Rillet Works, MCP Use Cases & Workflows, published April 11, 2026
Campfire: AI Integration with Model Context Protocol
Updated November 19, 2026: comparison refreshed against current vendor documentation.




