Buyer’s guide

How to choose an AI agent partner for finance operations

Choose an AI agent partner for finance operations by asking for evidence, not demos: an agent running in production against SAP, NetSuite or Dynamics, a clear human approval model, an evaluation and audit approach, who owns the platform and code, how handover works, and how fees tie to a measured business outcome. Start with one workflow.

A lot of finance and IT teams are on their second or third try at AI. The first round was a pilot that looked good in a demo. It ran on a copy of the data, with one person watching, and it never got permission to touch a live payable, receivable, reconciliation or order. So it stayed a pilot.

If you are shortlisting partners now, you are probably trying not to repeat that. This guide gives you 12 questions to put in an RFP or ask on a first call. It is written for CFOs, Controllers, VP Finance and CIO/VP IT on SAP, NetSuite or Dynamics, so the questions lean toward the ledger, approvals and audit, not generic software delivery. For each one: what a good answer sounds like, and what should worry you.

AI consulting firm, dev shop or agent partner: what are you buying?

Three kinds of firms will answer your RFP, and they sell different things.

AI consulting firm

Sells
Strategy: a roadmap, a use-case inventory, a maturity assessment.
Right when
You have no idea where to start.
Watch for
You already know which finance workflow hurts.

Dev shop

Sells
Engineering hours, built to your specification.
Right when
You can specify exactly what to build, and will run it yourself.
Watch for
The result is only as good as the spec, and someone has to run it afterward.

Agent partner

Sells
A working agent in production, on a platform, with the controls around it. The good ones also sell the exit: a handover so your team can run it.
Right when
You know the workflow and want it done, not documented.
Watch for
Ask for production evidence, not demos.

Know which one you need before you compare proposals. A strategy deck and a production agent are not substitutes.

The 12 questions

Put them in an RFP or ask them on a first call. For each one, what a good answer sounds like, and what should worry you.

Proof

Is it real, and will it work on your ERP?

01

Can you show an agent running in production against a system of record, not a demo?

Good answer

A walkthrough of a live agent (anonymized if needed) that reads from and writes to an ERP, with the approval step, the log, and a candid account of what goes wrong and how it gets caught. For an AP example, you should see an invoice matched to a PO and goods receipt, the proposed posting, the person who approved it, and a mismatch the agent flagged instead of posting.

Red flag

Slides, a sandbox, or a chatbot answering questions about a PDF. Also a demo where nothing ever fails.

02

Which ERP have you written back to, and through which controls?

Good answer

A specific system and release, the integration path, and how the agent's identity and permissions were set up. Writing back should go through the same authorization your team already uses, not around it.

Red flag

"we integrate with everything," with no detail on write-back. Reading from an ERP is the easy half.

03

Why does this workflow need an agent rather than simpler automation?

Good answer

A partner who is willing to say no. If a workflow is fully rule-based and the inputs are clean, a script or a workflow tool may do the job for less. Agents earn their place when inputs vary (invoices in any format, emails, customer POs) and judgment is involved.

Red flag

Every problem is an agent problem.

Control

Who approves, how it is tested, what the auditor sees.

04

Where does a person approve, and how does an agent earn its way out of approval?

Good answer

Every consequential action (releasing cash, committing spend, contacting a customer) stops at an approval gate at the start. The partner can explain the criteria for loosening a gate: a track record on a defined class of transactions, measured against an agreed threshold, with a person deciding. See approval gates.

Red flag

"fully autonomous" from day one, or no clear answer on who approves what.

05

How do you evaluate the agent before launch and after it is live?

Good answer

A test set built from your real transactions, including the awkward ones, run before launch and again every time a model, prompt or rule changes. After launch, ongoing monitoring and a person reviewing misses. Ask how they evaluate agents on live workflows, not just in testing.

Red flag

Accuracy promised before anyone has looked at your data.

06

What does the audit trail look like, and could our auditor read it?

Good answer

For any transaction, a record of what the agent read, what it proposed, why, who approved it and when. If an auditor pulls a vendor payment, your team should show the invoice, the PO and receipt match, the flags and the approver in one place.

Red flag

Logs that only engineers can read, or "we can pull that if you need it."

Ownership and data

Where your data goes, and what you keep.

07

Where does our data go, and do you train on it?

Good answer

A clear map of where data, embeddings and logs are stored and processed, ideally inside your environment. If they train models on your data, it happens only inside your environment, on models you own, and never feeds the partner's or a model provider's shared models. Frontier models see none of your data.

Red flag

Vague answers about "secure cloud," or terms that let your data improve the vendor's product or reach a model provider.

08

Who owns the platform, code, prompts and evaluation sets when you leave?

Good answer

You do. The platform runs in your environment. The code, prompts, configuration, evaluation sets and logs are yours, and the contract says so. Ask whether you can own the platform outright, not license it.

Red flag

A platform you cannot take with you, or ownership that covers "outputs" but not the logic that produced them.

09

What does handover look like, and what does our team need to run it?

Good answer

A documented handover with runbooks, evaluation harnesses and a defined operating model, plus a clear picture of who on your side will run it. Ideally the partner's engineers work inside your team during the build.

Red flag

No plan for handover, or a plan that amounts to a support contract with no end date.

Commercials

Who builds it, where to start, how fees work.

10

Who exactly will build it, and will we meet them before signing?

Good answer

Names. The engineers who will do the work join a call before you sign, and they can talk about your ERP and your finance workflow in detail.

Red flag

A senior team sells the project and a different team you never met delivers it.

11

How do you pick the first workflow, and what number will it move?

Good answer

A short, fixed-scope assessment that maps candidate workflows and ranks them with a number attached to each. The first workflow is one you already know is expensive. The metric and baseline are agreed before anything is built. Finance queues (AP, AR, reconciliations, close) and Ops Agents for inventory are both fair candidates when the cost is real.

Red flag

A long discovery phase with no defined output, or a first project picked because it makes a good demo.

12

How is your fee tied to that number, and what are run costs after launch?

Good answer

Some part of the fee depends on the agreed metric moving, with the metric, baseline and threshold written down before work starts. And a straight account of run costs after launch: model usage, infrastructure, monitoring and the people on your side.

Red flag

All fees fixed regardless of outcome, with no discussion of run costs until after launch.

Red flags when choosing an AI agent partner for finance operations

  • 01Guaranteed accuracy before discovery
  • 02"Fully autonomous" claims for cash, spend or customer contact
  • 03No approval gates
  • 04A platform you cannot take with you
  • 05Strategy decks with nothing shipped into an ERP

How to test a partner with a low-risk first step

You do not need to commit to a large program to test a partner.

  1. 01

    Pick one workflow

    One you already know is expensive, with a number you already watch.

  2. 02

    Ask for a fixed-scope assessment

    The workflow mapped, the environment audited, and a baseline for the number you want to move. It should be useful to you even if you stop there.

  3. 03

    Judge the output, not the pitch

    Decide on what the assessment produced. See how we deploy agents.

How Ferrata answers these 12 questions

We would rather you ask these questions of every partner, including us. Short version:

Production and ERP

Q1Q2

Our team comes from operating, not advising: $2.5B+ total GMV scaled across the platform and products we built, 120+ FTEs of capacity unlocked, and two recent deployments that started with SAP. More at operators, not an agency.

Approval

Q4

Every agent stops at an approval gate until it has earned its way out of one, and a gate comes off only when you decide the record supports it.

Data and training

Q7

The platform runs inside your environment, behind your firewall. Where it helps, we fine-tune open-weight models on your data there, for you only. Your data never leaves your environment and never reaches a frontier AI lab.

Ownership and handover

Q8Q9

You own the platform either way. Our AI Velocity POD, two forward-deployed engineers, builds inside your business and IT teams and trains them as it goes, so your team builds the next agent with the runbooks and evaluation harnesses documented.

First workflow

Q11

A 2-week assessment with a fixed scope, the platform in your environment in 3 weeks, and a first agent live in under 90 days. The Assess output is yours whether or not you continue.

Fees

Q12

The metric, the baseline and the threshold are agreed in the assessment, and if the number does not move you do not pay for the work that was meant to move it.

FAQ

What should I look for in an AI agent partner for finance operations?
Evidence of production agents on a system of record, a human approval model, an audit trail your controller can read, platform ownership after the partner leaves, a real handover plan and fees tied to a measured outcome. Start with one expensive workflow you already know.
What is the difference between an AI consulting firm and an AI agent partner?
Consulting firms sell roadmaps and assessments. Agent partners sell a working agent in production with controls and, ideally, a handover. Many buyers need both at different times. Do not price a deck against a live agent as if they were the same deliverable.
Who should own the IP and platform when a partner builds agents for us?
You should. The platform, code, prompts, evaluation sets and logs should run in your environment and remain yours after the partner leaves. If leaving means starting over from scratch on your own ERP, you are only renting your own operation.
What proof should an AI partner show before we sign?
A production walkthrough against an ERP (anonymized if needed), the approval and escalation path, the audit trail an auditor can read, named builders you can meet before contract, and a fixed-scope first step with a metric, baseline and threshold written down.
How long should it take to get a first AI agent into production?
Ask for a concrete plan. Ferrata's published engagement model puts the first agent live in under 90 days after Assess and Anchor, the stage that sets up the Brain. Treat any timeline that cannot name the first workflow, the ERP and the approval model as incomplete.
What should an AI agent development contract include?
Ownership of platform and artifacts, data terms (where any training happens, who owns the resulting models, and that your data never leaves your environment), approval and audit requirements, evaluation and acceptance criteria, clear handover obligations for your internal team, expected run costs after launch, and how fees relate to the agreed business metric.

Bring us one workflow. Bring these 12 questions to the call and we will answer every one.

Bring us one workflow.

Thirty minutes, no deck. Pick the process that costs you most in people, errors or delay, and we'll pressure-test whether an agent is genuinely the right answer for it.

  • An honest read on whether that workflow is a good agent candidate
  • What the first deployment would involve, and roughly what it costs
  • An honest answer if the sequencing is wrong and you should fix something else first
Book a discovery call