Ramp's Applied AI Solutions page lists four things its finance agents do. Three read like feature captions. The fourth is "Find the rule. Cite the source," and it gives the whole game away.
It looks like the smallest line on the page. It is actually the hardest problem in finance automation, written as if it were a checkbox.
Most of the coverage of the June 10 launch ran the expected story: Ramp is putting engineers next to enterprise finance teams to build agents for accounts payable, procurement, the close, and expense review. True enough. But "Find the rule. Cite the source" is not a description of a model. It is a description of a problem. Before an agent can do anything safe in finance, someone has to know which rule binds this decision, and where that rule is actually written down.
That is the work. The model is the easy part.
Walk one moment at the finance workbench
Picture a single invoice landing in accounts payable on a Tuesday.
It is from a vendor you have used for three years. The amount is eight percent over the last invoice. There is a master services agreement somewhere that caps annual price increases, but it was negotiated by someone who left in 2024. A PDF policy says invoices above a threshold need a second approver. The GL code the vendor expects is not the one your controller wants this quarter, because a department got reorganized. And there is a prior exception from last spring where finance let a similar overage through once, with a note that it was a one-time courtesy and not a precedent.
A generic AI agent can read all of that. It can summarize it cleanly. It can even produce a confident recommendation that looks right.
Looks right is the trap. Ramp says this directly on its own page: without the company's specific context, generic AI produces responses that look reliable but are not safe to act on. The invoice example is exactly where that breaks. The binding rule is not in the invoice. It is buried across a contract, a PDF, a chart of accounts, an approval matrix, and one human's memory of a decision made fourteen months ago.
To act on that invoice safely, the agent has to find the rule and cite the source. To do that, it needs a map.
The artifact is a buried-context map
The buried-context map is the operator-facing record of where the binding context for one finance decision actually lives.
It is not the invoice. It is not the model's answer. It is the thing that has to exist before either of those matters. Ramp's head of AI solutions, Ori Daniel, described the same shape in the launch release: "In finance, every decision depends on buried layers of context: the policy, the vendor, the contract, the approval chain, and the exception history." Name those layers, point each one at a real source, and you have a map an agent can read from and a reviewer can audit.
A useful map for one workflow has eight rows. The examples below are a running invoice scenario, not Ramp product claims.
| Map layer | What it answers | Invoice example |
|---|---|---|
| Policy source | Which written rule applies, and where is it? | AP policy PDF, section 4: second approver above $25k |
| Party identity | Who does the rule apply to here? | Vendor #3120, master services agreement on file |
| Contract term | Is there a negotiated exception or pricing logic? | MSA caps annual increases at five percent |
| GL mapping | Where does this decision land in accounting? | Code changed after Q1 reorg; controller wants new code |
| Approval chain | Who can approve, when, at what threshold? | AP lead under $25k; controller above; CFO above $100k |
| Exception history | Has a prior decision set or refused precedent? | 2025 overage allowed once, marked not a precedent |
| System of record | Which system holds the authoritative version? | ERP for GL, contract store for MSA, email for the 2025 note |
| Human review rule | When must the agent stop, cite, and ask? | Any term that conflicts with the contract cap |
Notice what is not on this list. There is no row for which model to use, how long the context window is, or how the prompt is phrased. Those are real engineering choices, but they are downstream. The map is what decides whether the agent's confident answer is grounded or just fluent.
Why the gap is a context gap, not a model gap
Ramp opens its pitch with a number from a Deloitte CFO survey, cited on the launch page: 87 percent of CFOs say AI is critical, and only 21 percent report measurable results. The usual read is that the technology is not ready. The buried-context map suggests a different reading.
Our read: the gap is less about who bought AI and more about who did the context-mapping work around it.
Finance is unusually hostile to ungrounded automation because its work spans systems and depends on judgment when exceptions appear. PYMNTS noted the same thing in its coverage of the launch: off-the-shelf tools struggle when a process runs across multiple systems, leans on company-specific policy, and needs a human call when something does not fit. A model that cannot point to the binding contract term is not a little less useful in that setting. It is unusable, because the whole job was knowing the term existed.
This is also why Ramp's own existing agents work where they work. On its Intelligence page, Ramp describes agents that code expenses, enforce policy, approve routine low-risk transactions, and escalate the ambiguous ones. Those agents are effective because Ramp already holds the context: it processes the spend, so it knows the vendor, the policy, and the pattern. Point the same model at a workflow where the context is scattered across a client's inherited ERP and a folder of PDFs, and nothing about the model changed. What changed is that the context it needs is no longer in reach.
What Ramp's method is really admitting
The most honest thing on the launch page is the method. Ramp is not selling a model you point at your data. It is sending engineers to sit with your finance team, map your data to how the business actually runs, and only then build agents that read and write into your existing systems. The launch lists the steps plainly: context extraction, workflow prioritization, system integration, agent design, human review and controls, and continuous improvement.
Two of those steps come before anyone builds an agent. Context extraction and workflow prioritization are the map. Ramp is charging for embedded engineering time because the buried context will not extract itself, and because the company that processes $200 billion a year for tens of thousands of businesses has learned that the map is the deliverable. The agent is what you get after the map exists.
You do not need Ramp's engineers to take the lesson. You need the map.
Map one workflow before an agent gets write access
If you are a controller or a finance ops lead being asked to put an agent into AP, procurement, the close, or expense review, the useful first move is not a vendor demo. It is to pick one workflow and fill in the eight rows for a single real case.
Start narrow. One invoice type, one expense category, one reconciliation. For that one case, write down the binding policy and where it lives, the party it applies to, the contract term that could override the default, the GL mapping, the approval chain and thresholds, the relevant exception history, the system of record for each fact, and the exact point where the agent must stop and ask a person.
You will learn two things fast. First, how much of the context is currently in someone's head rather than a system anyone could cite. Second, where the binding rule and the system of record disagree, which is the failure that produces a clean-looking answer that is quietly wrong.
Both of those are cheaper to find on paper than in production. A finance agent that can read everything and cite nothing is the expensive version of the same problem. This is the same instinct behind keeping a state ledger before you add more context, and the same reason a back-office workflow benefits from a clear map before automation: the agent is only as safe as the record it can point back to.
When you do wire an agent in, the map becomes the control surface. It tells the agent what it may read, what it may write, and where it must route a decision to a person. That boundary is what an approval queue and a set of AI workflow controls are for. It is also where process automation does the real work, because mapping one workflow honestly is what makes the next ten safe to wire up.
Ramp put the whole argument in four words on a product page. Find the rule. Cite the source. The rule is buried, the source is scattered, and the map is the work. Build it before the agent gets write access.
Map the buried context in one finance workflow before an agent gets write access. Start with one workflow.
Implementation help
Map one finance workflow before an agent gets write access
BaristaLabs helps finance and ops teams turn buried context into a reviewable map: the binding policy, the parties, the contract terms, the GL path, the approval chain, the exception history, and the stop conditions an agent must cite.
Start with one workflow lane before giving a finance agent broader autonomy.
Practical AI Workflow Notes
Want more practical AI operations ideas?
Get short notes on applying AI inside real small-business workflows — from document handling and customer follow-up to internal reporting, compliance, and automation guardrails.
Share this post