A supplier invoice arrives with two plausible purchase order numbers on it. One step in your process copies the header fields into the ledger. Another asks someone to work out which PO applies. A third asks a buyer to approve a price that changed since the order. Those are three different jobs, and the right answer to “bot, AI agent, or outsourced team?” can be different for each one.
Compare steps, not whole products. Decide who or what does each step, who owns the decision, and what happens when it fails.
What each option actually is
Robotic process automation (RPA) runs configured steps in an application: open this screen, copy this field, click this button, export this report. It does the same thing every time. Microsoft’s documentation for Power Automate desktop flows, for example, describes how flows interact with applications and websites through UI elements, the buttons and fields on the screen (Microsoft’s guide to automating with UI elements). When the screen is stable, the bot is fast and predictable. When a field moves or an unexpected dialog appears, the flow has to stop or it does the wrong thing.
An AI agent uses a model to interpret text and propose an action: read an invoice with an unfamiliar layout, suggest which open PO it relates to, draft the coding, and route anything uncertain to a reviewer. It handles variation better than a scripted bot. It also needs explicit limits on what it can read, prepare, and write.
Outsourced accounting puts people at a service provider on the work. They can make phone calls, chase a buyer, and exercise judgment if the engagement covers it. They work inside your systems under access you grant, and their scope is whatever the service agreement says.
The three can coexist in one workflow.
Assign the tool to the step
Take one workflow, say non-PO and PO vendor bills, and list its steps. For each one, ask what kind of work it is.
| Step | Nature of the work | Usually fits | What to verify |
|---|---|---|---|
| Export the open-PO report from a stable screen | Fixed, repetitive | RPA or a scheduled report | Period, row count, and file before anyone uses it |
| Read invoices with varying layouts | Interpretation | AI agent | Source line behind each extracted field; uncertain values go to review |
| Compare invoice and PO quantities | Rule-based comparison | ERP matching or explicit rules | Approved unit conversion and tolerance |
| Investigate a disputed line with the supplier | Judgment plus phone or email | Your staff or an outsourced team | Who has authority to agree a credit |
| Approve a supplier price change | Authority | The buyer | Old and new price, with the decision recorded |
| Create the approved bill in the ledger | Controlled write | Supported integration path | Saved record reconciles to the approval |
| Release payment | Authority | Named person under your policy | Separation from whoever prepared the bill |
Two rows in that table never move to a tool or a provider by default: approving exceptions and releasing payment. You can delegate the work around them. The authority needs a named owner.
How each option fails
A demo shows the clean case. Your decision should rest on the failure cases, because that is where staff time goes.
| Failure | RPA | AI agent | Outsourced team |
|---|---|---|---|
| Screen or layout changes | Flow breaks or clicks the wrong field unless it is built to stop | Usually adapts to a new invoice layout; an interface change still needs testing | Staff adapt, but may follow an outdated procedure |
| Ambiguous document (two PO numbers) | Takes whatever is in the configured position | Should present both candidates with source text and leave the item open | Investigates if in scope; otherwise sends it back |
| Duplicate invoice resent after posting | Creates a second bill unless a lookup step exists | Should find the earlier bill before writing | Depends on the provider’s checks and your system’s duplicate controls |
| Write times out | May retry blindly | Should check whether the record exists before retrying | Person checks, if trained to |
| Policy change | Someone reconfigures the flow | Someone updates rules and instructions | Someone updates the engagement and procedures |
The “should” entries in the AI column are things to test, not assumptions. Plenty of tools will happily pick one PO number and move on.
Run the same three invoices through each option
Build a small test set from records your team already processed, so the right answers are known:
- A clean bill. Verify the ledger record afterward: vendor, entity, date, amount, coding, attached document.
- A PDF with two plausible PO numbers. Check whether the reviewer receives both numbers and the source text, or whether one was silently chosen.
- An invoice resent after it was already posted. Check whether the workflow finds the earlier bill before attempting another write.
Then change one thing. Move a field on the test screen, or use a new supplier layout. Record what stopped, what produced a wrong result, and how long it took someone to repair it. A bot that stops loudly is better than one that keeps running and posts to the wrong account.
For an outsourced provider, hand over the disputed bill and ask which steps come back to your staff.
Ask the questions that decide ownership
When you compare an outsourced service with software, use one transaction and put the same questions to both.
| Question | Ask the accounting service | Ask the software vendor |
|---|---|---|
| Who prepares the invoice draft? | Who enters it and checks entity and coding? | Which fields are proposed, and where can AP see their source? |
| Who investigates a discrepancy? | Who contacts the buyer or supplier, and does the engagement cover it? | What evidence goes to the reviewer, and who follows up? |
| Who records the approved result? | Who posts the entry and verifies the saved record? | What can the connection write, and how does it handle a failed attempt? |
| Who resolves a mistake? | Who corrects the record and reports the impact? | How does your team find affected records and stop further writes? |
If you need a named person to own a close task or apply accounting judgment, put that responsibility in the service agreement. If you already have a reviewer and the bottleneck is gathering evidence, software on that step is the more direct test.
Controls at a service provider
Decide who approves exceptions, verifies bank detail changes, and releases payments, and keep that list in your own policy, not the provider’s. Ask which controls the provider operates and which stay with you. If they offer a SOC 1 report, check that it covers the service you are buying. The AICPA describes SOC 1 engagements as examinations of service organization controls relevant to user entities’ internal control over financial reporting (AICPA SOC 1 guide). A report on a different service line tells you little about yours.
Read, prepare, write: levels of agent authority
Agents inside or connected to an ERP vary widely in what they can do. Separate four levels:
| Level | What to verify |
|---|---|
| Read | The answer names the records and the period, entity, and currency it used |
| Prepare | The proposed change shows current and new values and waits for review |
| Write | The role has authority for that action and the saved record matches the approval |
| Recover | A reviewer can see what succeeded, what failed, and what needs correction |
The available actions depend on the tools and roles you enable. Oracle’s NetSuite AI Connector Service FAQ, for instance, says the SuiteQL tools are read-only and that queries respect the permissions of the NetSuite role (Oracle’s NetSuite AI Connector Service FAQ). Ask any vendor to show the specific tool and role behind each action it claims.
Test it on one blocked bill: ask why it is blocked and check the invoice, PO, and receipt cited, then ask for the exception prepared for purchasing. Enable a write only after both hold up. Payment release stays where your policy puts it.
Compare cost per correctly completed item
License fees are the easy part. The real comparison is total cost divided by the transactions completed correctly, including review, exception handling, maintenance, and correction time. The AP automation pricing and ROI guide covers how to build that model in detail.
Worked example (illustrative)
These figures are placeholders to show the arithmetic, not market prices. Replace every input with your own quotes and timings. Assume 1,000 vendor invoices a month and an internal staff cost of $40 per hour.
| Input | RPA | AI agent | Outsourced |
|---|---|---|---|
| Monthly fee | $1,500 | $2,500 | $3.00 × 1,000 = $3,000 |
| Items needing exception handling | 250 at 12 min = 50 h | 120 at 10 min = 20 h | 150 returned to you at 8 min = 20 h |
| Routine review by your staff | 750 at 1 min = 12.5 h | 1,000 drafts at 1.5 min = 25 h | Oversight sampling = 6 h |
| Maintenance / rule updates | 10 h | 5 h | Included in fee |
| Internal hours | 72.5 h | 50 h | 26 h |
| Internal cost at $40/h | $2,900 | $2,000 | $1,040 |
| Total monthly cost | $4,400 | $4,500 | $4,040 |
| Cost per invoice | $4.40 | $4.50 | $4.04 |
On these inputs, the three options land within 46 cents of each other per invoice, and the exception rate is the input that moves the answer most. If the RPA exception rate dropped from 25% to 10%, exceptions would fall to 100 × 12 min = 20 h, routine review would rise to 900 × 1 min = 15 h, and with 10 h of maintenance the internal hours would be 45 h, or $1,800. Total cost becomes $3,300, or $3.30 per invoice.
That is why you run the failure tests first. The exception rate on your own invoices, not a vendor’s average, decides the answer. Also count errors that reached the ledger separately: one wrong posting found at close can cost more time than a week of routine review.
What changes for your team’s roles
The U.S. Bureau of Labor Statistics expects software to automate routine bookkeeping tasks and shift some of the remaining work toward analysis (BLS Occupational Outlook Handbook). That is an occupation-wide trend. For your own team, look at one role’s desk and divide it:
| Work | What software or a provider can prepare | What your reviewer still owns |
|---|---|---|
| Vendor bill | Dates, amounts, suggested coding, invoice attached | Whether a charge belongs in this period (a 12-month prepayment is not last month’s expense) |
| Customer receipt | Candidate open invoices for the remittance | An unexplained short payment before granting credit |
| Bank reconciliation | Grouped candidate matches with references | Which unmatched items need investigation; sign-off |
| Close schedule | Gathered transactions and a flag on a changed balance | The explanation and any adjustment |
Then write down who owns the remaining work and who covers it at close. The guide to scaling finance without hiring includes an absence drill that tests whether a backup can work from the same evidence.
When to keep what you have
- Keep an existing bot if it handles its step reliably and someone on the team can maintain it.
- Add an AI agent to the step where interpreting inconsistent documents creates most of the work, and keep the existing checks that protect the ledger.
- Use an outsourced team where the work needs people to call, chase, and decide, and where you can write the scope and ownership down.
- Fix ownership first if the remaining delay is an approval nobody owns. No tool clears that.
For the audit trail each option should leave, see audit trails and controls for automated AP and AR. For structuring a vendor evaluation, the finance automation buying guide has a scoring table.
Before you sign with anyone
- Name the owner for exception approval, bank detail changes, and payment release.
- Run the three-invoice test set through every option on your shortlist.
- Recalculate cost per correctly completed item with the exception rate you measured.