Operations · 7 min read · Aug 15, 2026

Finance pays back last. That is not a failure.

The reported results from back-office agents are strong enough to look implausible, and the reported payback periods are long enough to look like a contradiction. They are not. The gap between them tells you which process to automate first.

What is actually being reported

Organisations deploying back-office agents report 70–90% reductions in invoice processing time, and automated invoicing, forecasting and expense auditing accelerating close processes by 30–50%. In HR, time-to-hire is reported to fall by around half. In financial services, back-office processing — invoice management, disputes, compliance monitoring — is among the largest deployment categories.

Vendor numbers point the same way. Salesforce reported Agentforce reaching $800M ARR in Q4 FY2026, up 169% year on year across 29,000 customer deployments, with customers claiming over $100M in annualised cost savings; in April 2026 it extended into back-office work explicitly with Agentforce Operations.

And yet finance pays back last

Median months to payback for agent deployments: sales development 3.4 months, median across all functions 5.1 months, finance and operations 8.9 months.
Median time to measured payback. Finance and operations take roughly two and a half times as long as sales development.

Median time-to-value across functions is 5.1 months. Sales development agents pay back in 3.4. Finance and operations agents take 8.9 — nearly triple the SDR figure, against a process with larger reported efficiency gains. The apparent contradiction resolves once you look at what the two are allowed to do.

An SDR agent that writes a poor email costs one wasted send. An invoice-coding agent that miscodes costs a restatement, an audit finding, or a payment that should not have left. Where output touches an audit trail, a regulated process or a contractual obligation, organisations keep more humans in the loop — deliberately. The longer payback is the cost of that choice, and it is usually the right one.

The finance payback gap is not evidence that finance agents work less well. It is evidence that finance is where being wrong is most expensive, and organisations are pricing that correctly.

The same logic explains the sector spread

Banking and insurance run near 47% adoption of at least one production agent; healthcare sits near 18% and government near 14%. Banking has both the volume and mature control frameworks to absorb an agent inside existing checks. Healthcare and government have the volume but far less tolerance for an unexplainable output. The constraint is not appetite or capability — it is what the surrounding process can verify.

Sequence by checkability, not by size of the prize

The instinct is to start where the manual effort is greatest. The better rule is to start where the output is easiest to check, because that is what determines whether the thing ever gets trusted enough to run. A useful ordering for a multi-entity group:

ProcessWhy it comes early or lateCheck available
Supplier invoice captureEarly — ground truth exists on the documentThe invoice itself
Item and supplier normalisationEarly — errors are visible and reversibleHuman review queue
Variance flagging on closeMiddle — flags a human, does not decideThe close report
Expense policy checksMiddle — clear rules, contested edgesThe policy
Journal posting, payment releaseLate — irreversible and auditedApproval chain only

Note what the last row has in common with the long payback: the check is a person, not a document. That is precisely why it takes longer to earn back, and why attempting it first is how a programme loses its mandate.

Build or buy

The pattern across 2026 deployments is consistent: commodity back-office tasks are increasingly bought as vertical or platform-maintained agents, while genuinely differentiating capability is built. The test is whether the process is one your competitors run identically. Invoice OCR is not a competitive advantage. Knowing that two of your outlets pay different prices to the same supplier — because you did the item normalisation nobody else bothered with — is.

That work is specific to your estate, which is why it does not come in a box. The rest of the method is in consolidating reporting across different POS systems, and there are more working notes in Ontilus's insights.

Sources

  • BCG and Forrester surveys, 2026 — median 5.1 months to value; SDR 3.4; finance and operations 8.9.
  • 2026 enterprise deployment reporting — 70–90% reduction in invoice processing time; 30–50% faster close.
  • S&P Global Market Intelligence and McKinsey, 2026 — sector adoption: banking and insurance 47%, healthcare 18%, government 14%.
  • Salesforce — Agentforce $800M ARR Q4 FY2026, 29,000 deployments; Agentforce Operations launched April 2026.

Dealing with this in your own group?

We answer scoping questions before there's a contract in sight — including the ones about cost and data handling.

Questions

Short answers,
in full.

The questions this article gets asked most, answered so each one stands on its own.

Talk to us
How long do back-office AI agents take to pay back?

Survey data from 2026 puts median time-to-value across functions at 5.1 months, with finance and operations agents at 8.9 months and sales development agents at 3.4. Finance takes longest because its outputs touch audit trails, regulated processes and contractual obligations, so organisations deliberately keep more human review in the loop.

Which finance process should we automate first?

The one whose output is easiest to check, not the one consuming the most hours. Supplier invoice capture is a strong first candidate because the ground truth sits on the document itself. Journal posting and payment release come last: they are irreversible and audited, and their only check is an approval chain rather than a document.

Are the reported efficiency gains from agents real?

The reported figures — 70–90% reductions in invoice processing time, close cycles 30–50% faster — come from organisations that reached production, which is a minority of those that started. They describe what works after deployment, not the probability of getting there. Both numbers are worth reading together.

Should we build back-office agents or buy them?

Buy where the process is identical to your competitors' — invoice OCR and standard document extraction are commodity. Build where the advantage is specific to your estate, such as normalising item and supplier names across outlets so cross-outlet price comparison becomes possible. That work depends on your own data and does not arrive in a product.