Resources ยท Guide
Accounts Payable Automation: An Evaluation Guide
What to look at, what to ask and what to measure when choosing how to automate accounts payable.
Most AP automation evaluations compare feature lists. The lists look alike, so the decision ends up resting on the demo, and the demo runs on the vendor's invoices. This guide is for evaluating on yours. Map the process as it runs today, count where it stops, write requirements from those counts, test every vendor on the same pack, and roll out in stages that each have a test you can fail.
Start with the process you really run
Write the process down with the people who run it. The policy document describes the process someone intended; the clerks can describe the one that runs. Keep it to ten steps or fewer, each with its input, its output and who does it. An honest map of a payables process includes steps the policy never mentions: the invoice forwarded from a buyer's inbox, the spreadsheet where someone tracks what is waiting for approval, the phone call to stores to ask whether the goods arrived.
Then add four numbers for one recent month. Invoices received, split into PO-backed and non-PO. The share that arrived as scans or photos. The number of distinct suppliers. The number of entities or GSTINs you are invoiced against. These decide more about the project than any feature list does. A queue that is mostly non-PO service invoices needs good coding and routing; one that is mostly PO-backed goods needs good matching.
Mark where each piece of evidence lives. Purchase orders in the ERP, goods receipts in a stores register, invoices in a mailbox, approvals in email. Every hop between systems is a place the automation needs an integration, a file or a person, and each hop belongs in the requirements.
Count the exceptions before you choose
For one month, log every invoice that needed a person for any reason other than routine approval, and write the reason down each time. Group the reasons. Do this before you see a single demo, because the groups tell you which kind of tool you need.
One month of AP exceptions, counted by cause · Illustrative example
Illustrative data. Count your own month before choosing a tool: if most holds come from policy and master data, better extraction will not shorten the queue.
In the example, reading the document caused about one exception in ten. The rest came from a price outside tolerance, a missing PO, a goods receipt not yet posted and a vendor missing from the master. A tool that reads invoices beautifully will not shorten that queue on its own. What shortens it is matching with your tolerances, a rule for the non-PO case, a vendor master someone owns, and a workflow that waits for a GRN instead of failing on it. If your count looks like this one, weight the evaluation towards matching and rules. If reading dominates because your suppliers send scans and handwritten amendments, weight it towards extraction and review.
Requirements, stage by stage
Write each requirement as something a vendor can demonstrate on your documents. "Supports three-way matching" is a checkbox. "Matches this invoice to this PO and these two partial GRNs, with our pack-size conversion" is a test. The table is a starting set; replace the third column with examples from your own exception log.
| Stage | Require | Ask to see |
|---|---|---|
| Intake | Every channel you use (mailbox, portal, scan folder, API), logged on arrival | The arrival time on an invoice that was forwarded twice before it reached AP |
| Extraction | Header, every line and the tax block, with a confidence score per field | A forty-line invoice with a stamp across the totals, and what happens to the field it could not read |
| Validation | Vendor master match, GSTIN status, bank details, duplicates including near duplicates | The same invoice sent by email and by portal, one copy with a changed prefix |
| Matching | Two- and three-way by default, six-way where the data exists; your tolerances, unit conversions, partial receipts | A short delivery matched to what was received, with the balance left open |
| Non-PO invoices | Coding to account and cost centre, routing to the budget owner | A utility bill and a software subscription, coded and routed |
| Approvals | Bands by amount and category, delegation, escalation, a recorded reason | An approval from a phone, and what happens while the approver is on leave |
| Posting | Your accounts and dimensions, the document attached, the ERP number returned | The entry in your own ERP test instance |
| Payment status | Scheduling on terms, status written back from the bank file | An invoice marked paid because the bank file said so |
| Audit | Per invoice: what was read, which rules ran, who decided, what posted | The complete record for one invoice, exported |
Integration deserves its own line. Ask for the method per system: native, API, file or custom. A vendor who says they integrate with your ERP without naming the method has not scoped it yet, and the method usually decides the timeline. FinSuite Integrations shows what moves in a FinSuite deployment and how. The GST checks in the validation row are covered in detail in the GST automation guide.
Controls an auditor will test
Automation moves the checks earlier in the process. Confirm that it drops none of them on the way. Ask each vendor to show these working, on your data where possible:
- Segregation of duties. The person who corrects a field or codes an invoice cannot approve it, and the approver cannot release the payment. The system enforces the roles.
- Thresholds held as recorded configuration. Confidence thresholds, match tolerances and approval bands each have an owner, and every change records who made it and when.
- Bank detail changes verified out of band. A changed account is held until someone confirms it through a channel other than the email that announced it.
- Duplicates stopped before payment, near duplicates included: a changed prefix, a resent scan, the same invoice through two channels.
- A per-invoice trail that exports. What was read, which rules ran, who decided and what posted, in a form your internal and statutory auditors can use.
- Data handling you can state. Where invoices and bank details are held, who can see them, and a retention period set by your policy.
Review is one of these controls, and the one most often designed badly. The human-in-the-loop guide covers how to set triggers, queues and thresholds so that review keeps working as volume grows.
How to run the evaluation
Give every shortlisted vendor the same pack: fifty to a hundred invoices from last month as they arrived, the matching POs and GRNs, an extract of your vendor master, your approval policy, and the list of exceptions you logged. Keep the answer key: what your team did with each invoice, and why.
Ask each vendor to return, per invoice, what was read, what was matched, what was held and for what reason. Then compare against your key. The most expensive result is an invoice the tool would have posted that your key says should have been held. Next come holds for reasons that are not reasons, because each one costs a person's time. Last, note how long each vendor took to return results on a pack they had never seen. It is a fair preview of onboarding.
Be wary of a demo that only uses the vendor's own invoices, of accuracy claims that do not say what was counted, and of any answer to a what-happens-when question that starts with the system learning over time. The IDP evaluation guide goes deeper on testing the extraction step by itself.
A rollout plan that survives the first month
Five stages, each with an exit test
The shadow run is the cheapest place to learn that a tool mishandles your credit notes. Skip it and you learn the same thing in production.
Start narrow. One entity, or one group of suppliers that covers a large share of volume, with the PO-backed flow first because the match gives the clearest early signal. Add the non-PO flow and the other entities when the exception causes are falling, and let those counts set the date.
The shadow run is the stage teams most want to skip, and the one most worth keeping. A month of your real invoices goes through capture, rules and matching with nothing posted, and the result is compared line by line with what your team did. It surfaces the credit notes the tool mishandles, the supplier who prints the PO number in the remarks line and the approval rule nobody wrote down, while fixing them still costs nothing.
Going live on a single AP workflow typically takes 4 to 12 weeks, an indicative range that varies by document types, workflows, integrations and deployment. Expect the early weeks to go on your documents and your rules as they really are.
What to measure, and the baseline
Take the baseline before the pilot, from the same sources you will read afterwards. Then read the same measures every week of the pilot.
What to measure in the pilot, and how to count it
Take every baseline before the pilot starts, from the same sources you will read afterwards. A measure with no before-number is an opinion.
Two cautions. Report by invoice class, because PO invoices, non-PO invoices and credit notes move at different speeds and an average hides the difference. And treat touches per invoice as the headline but exceptions by cause as the diagnosis: a falling touch rate alongside a rising share of one cause tells you exactly where to work next. What these measures should look like once the workflow is running is set out on the Accounts Payable Automation page.
What we would not automate
A disputed invoice. An approval above the band your policy sets. A change of vendor bank details. The accounting treatment of something genuinely unusual. A system should hold each of these with the evidence attached and a named owner, and record that it waited. Automating them would shorten the queue and weaken the control, which is the wrong trade for a payables team.
How to test this on your own invoices
- Map the process in ten steps with the clerks who run it.
- Log a month of exceptions by cause, with an owner for each cause.
- Pull fifty to a hundred invoices as they arrived, with their POs, GRNs and the outcome your team reached.
- Write the requirements table with your own examples in the third column.
- Give every vendor the same pack, and score the results against your key.
Whichever vendor you choose, that pack is the evaluation. If you want to run it with us, the finance workflow assessment scores the process first, and the DocPro document test reads one of your invoices now.
Where PTAS AI fits
Each stage of the evaluation, and the page that covers it
FAQ
Questions finance teams ask when evaluating AP automation
How is this guide different from the Accounts Payable Automation page?
The solution page describes the automated payables workflow and what changes once it runs. This guide covers the decision before that: mapping your current process, writing requirements, testing vendors on your own invoices and rolling out without losing control. It applies whichever vendor you choose.
How many invoices should the evaluation pack contain?
Fifty to a hundred from one recent month, as they arrived, is usually enough to show how a tool handles your layouts, your exceptions and your credit notes. Include the awkward ones: scans, multi-page invoices, non-PO bills and the duplicates you caught. A pack of clean PDFs makes every vendor look the same.
Should we clean the vendor master first?
If your exception count shows many vendors missing from the master or duplicated in it, yes, before or alongside the pilot. Matching and vendor checks depend on the master, and automation run on a poor one produces confident errors. The work is tedious and it pays back in every later step.
Do we need a three-way match for every invoice?
No. Services and subscriptions usually have no goods receipt, so they match two ways or are coded without a PO. Require the tool to handle all three cases, and start the pilot with the PO-backed goods flow, where the match gives the clearest signal.
Is six-way matching worth asking for?
Ask for it, but do not start with it. Six-way matching adds the vendor's identity, the contract terms and the tax, bank and duplicate checks to the three-way match, and those catch the costliest errors: a changed bank account, an invoice resubmitted under a new number. Each check needs data behind it, so pilot on three-way and switch the rest on as the vendor master and contracts are in shape.
When will we know whether it is working?
The shadow run gives the first real answer, because it compares the tool's decisions with your team's on a month of real invoices before anything posts. Going live on a single workflow typically takes 4 to 12 weeks, an indicative range that varies by document types, workflows, integrations and deployment.
Keep reading
Related reading
Accounts Payable Automation
The automated payables workflow end to end, and what changes when it runs.
Read more →FinSuite AP Automation
The module that runs matching, approvals and the payables queue.
Read more →How to Evaluate Intelligent Document Processing
A test plan for the extraction step, run on your own documents.
Read more →GST Automation Guide
The tax checks that sit inside AP validation, line by line.
Read more →Next step
Score your payables process before you choose a tool
Bring a month of invoices and your exception log. We will score the process with you, show where the queue really stops and suggest the flow to pilot first.