Will AI Agents Be Standard in Finance Operations by 2031?

The question for a leadership team right now is whether to wait for agents to mature or start now. Nobody can say for sure what 2031 looks like. This article sets out how fast things have moved since 2021, what the main forecasters say, and which payments and finance processes ScalePoint Research expects to run on agents by 2031, using its process map.

Five years back

In 2021, automation in finance operations meant rules engines and RPA bots. AI models could write fluent text, but firms didn't use them on ledgers or payments. The changes since then came in a few steps.

Function calling and MCP are what turned models into agents. Function calling lets a model ask a piece of code to run, such as "search the ledger for this reference". MCP is a common way for a system to offer those tools. A vendor can publish one MCP connection to its data, and any model that supports MCP can then use it, so each firm doesn't have to write its own integration. MCP isn't tied to Anthropic's models, and since December 2025 it has been run by the Agentic AI Foundation, part of the Linux Foundation. The guide to models, copilots and agents explains these terms.

How fast it's moving

Two measures give a sense of the pace.

The first is how long a task a model can complete. METR, a research group, measures this as the time a skilled person would need for tasks a model gets right about half the time. GPT-4, in March 2023, managed tasks of about 4 minutes. Claude Opus 4.6, in February 2026, managed about 12 hours. That length has doubled roughly every four months since 2023. Two caveats: the tasks are software tasks, and at an 80% success rate the lengths are several times shorter. Finance teams need far better than 80% before letting an agent near a ledger, which is why the approval step matters.

The second is price. Epoch AI found that the price of reaching a given level of model performance fell by between 9 and 900 times a year, depending on the task. Matching GPT-4 on a set of hard science questions got about 40 times cheaper each year. An agent that's too expensive per case today may not be in two years.

On adoption, the Bank of England and FCA survey from 2024 found that 75% of UK financial services firms already used AI in some form. Only 2% of their AI use cases were fully autonomous, and that low share fits the rest of this article. Finance firms keep a person on the decisions that matter by design, and an agent that works the case while a person approves the result counts as semi-autonomous in that survey. Hence the number to watch is how much of the work AI does with a person approving.

What the forecasters say

• Gartner, June 2025: by 2028, 15% of day-to-day work decisions will be made by AI agents on their own, and a third of enterprise software applications will include agents. Over 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear value or weak risk controls.
• Gartner, 2026: by 2029, 30% of finance teams will have maximised AI-driven efficiency in their transactional processes.

So Gartner expects agents to spread quickly and a large share of early projects to be cancelled at the same time. Neither forecast says which processes go first. The next section uses the process map for that.

Which processes go agentic by 2031

The ScalePoint Research process map covers 51 payments and finance processes across six types of firm, from merchants and platforms to acquirers, banks, EMIs and stablecoin firms. The processes sit in 11 clusters that share their steps and their vendors, and each process is rated for how well it suits an agent:

• Agent workflow (22 processes). The work comes as a queue of cases, each case needs several lookups across systems, and a person can check the result against the records.
• Agent-assisted (14 processes). An agent can gather the facts and draft, and a person makes the decision.
• Single step (15 processes). The work is one read or one calculation, or it already runs on a billing or pricing engine. A model may do the step, and there's no case for an agent to work.

Here is how the 11 clusters split, and ScalePoint Research's view of where each one will be by 2031.

Case work goes first

Reconciliation, disputes, payment exceptions and alert review hold 18 of the 22 processes rated agent workflow. A settlement break, a chargeback, a payment rejected for a missing field and a sanctions hit all work the same way. There are many similar cases, each one needs several systems, and a person can check the result. Agents for all four clusters are already on sale, and the cost per case keeps falling. ScalePoint Research expects an agent working the queue, with people approving, to be the normal set-up for these processes well before 2031.

Onboarding is close behind. Merchant onboarding, business KYB, periodic KYC reviews and sub-merchant onboarding are rated agent workflow, because building the file means pulling registry data, documents and ownership records from several places. Approving the customer stays with a person. Consumer KYC is already automated by specialist vendors, so an agent there only works the referrals.

An agent prepares, a person decides

In 14 processes on the map, the agent gathers the facts and drafts and a person makes the call. Examples are filing a suspicious activity report (SAR), setting a merchant's reserves, approving a deal price, signing off a regulatory return and posting month-end journals. The agent still saves most of the time, because gathering the facts is most of the work. The line on who decides will move slowly, and in some cases not at all, because it's set by regulation or by the firm's own sign-off rules. The next section covers which rules.

Single steps, done by a model at volume

Fifteen processes on the map are single steps. Examples are reading a statement, checking a contract against billing, reviewing FX margins against the price list, positioning cash and analysing declines. Some, such as merchant fee billing, FX quotes and payment routing, already run on billing or pricing engines. ScalePoint Research doesn't expect these processes to move to agents by 2031. A model may still do the step, called once per document or deal, and often thousands of times a month. Picking the model on accuracy and cost per call is then the main decision, and that's what the ScalePoint Research tests measure. In Test 1, models costing 1 to 3 cents a statement came close to the best at reading card acquiring statements.

Acquirer deal pricing and FX margin review are low-volume analysis jobs on a firm's own data, run by small commercial or finance teams. A firm that wants AI help here is more likely to build a small tool of its own, with a model doing one step, than to wait for a vendor product.

Evidence so far

Each ScalePoint Research test adds a row to this table, with what that test shows for the point of this article.

Which rules keep a person in the decision

Some decisions stay with a person because a regulation says so. Others stay with a person because the firm's own sign-off rules say so, and the firm can change those itself.

• Reporting suspected money laundering. In the UK, the Proceeds of Crime Act 2002 gives the decision to report to the National Crime Agency to the firm's nominated officer, usually the MLRO. EU and US anti-money laundering rules also put reporting with a named compliance officer. An agent can draft the SAR, and a person decides whether to file it.
• Decisions about a customer made by automated means alone. GDPR in the EU limits decisions with legal or similarly significant effects made about a person by automated means alone, such as refusing to onboard a customer. The UK version, amended in 2025, still gives the customer the right to have a person review the decision.
• Consumer credit decisions. The EU AI Act classes AI used to assess the creditworthiness of individuals as high-risk, with requirements for human oversight. After the EU's 2026 changes to the Act, those requirements are due to apply from December 2027.
• Accountability. In the UK, the FCA relies on the Senior Managers and Certification Regime and the Consumer Duty, so the senior manager for a function stays accountable for what an agent does in it. In the US, the Federal Reserve and the OCC replaced their model risk guidance, SR 11-7, in April 2026. The new guidance leaves generative and agentic AI out of scope for now, and banks stay accountable for the results under their existing controls and third-party risk guidance.
• The firm's own sign-off rules. Setting a merchant's reserves or approving a deal price sits with a credit or pricing committee because the firm's delegated authorities say so. No regulation requires a person here, so the firm can decide to let an agent approve small cases within limits.

The article on buying operations software with AI inside covers DORA and the EU AI Act from the buyer's side.

Who builds and who buys

Large banks are building their own. BNY runs its own AI platform, called Eliza, partly on Microsoft Azure, and designed it to work with models from different makers. BNY has nearly 140 "digital employees" running on Eliza. According to Microsoft's write-up, the first one in payments cut the time to validate a payment from five or six minutes to under 30 seconds. Hence Eliza is an in-house build on bought cloud and bought models.

Firms without that engineering capacity buy agents from vendors. Gradient Labs, a London vendor of agents for customer operations, disputes and compliance, lists UK fintechs such as Plum and Pockit among its customers. Either way, the agent needs to reach the firm's data.

Data access sets the pace

An agent can only use the systems its tools can reach. A firm whose settlement files, ledger and contracts sit in systems with APIs can give an agent read-only access today. A firm whose settlement files arrive by email and whose contracts are PDFs in a shared drive has groundwork to do first, and that work is likely to set its timing more than the models will:

• getting the files out of inboxes and into one store the agent's tools can search
• putting the key contract terms, such as fee rates and time limits, into a table the tools can read
• cleaning up reference data, such as merchant IDs and account codes, so records from different systems can be matched
• setting up read-only access, permissions and logging for each tool

What to do in 2027

• Pick one high-volume queue, such as reconciliation breaks or chargebacks, and measure it now: cases per month, hours per case, money recovered. That's the baseline any agent has to beat.
• Get the data for that queue reachable: settlement files, ledger and contracts in one place, ideally behind an API.
• Ask every vendor which model they use, what it costs per case at your volumes, and whether you can switch models later.
• Decide where the approval points sit before you sign a contract.
• For single-step work such as reading statements, test a few models on your own documents before choosing one, because price is a poor guide to accuracy.

Sources

• METR, Measuring AI Ability to Complete Long Tasks, March 2025 paper and 2026 updates
• Epoch AI, LLM inference price trends, 12 March 2025
• FCA, AI in UK financial services, 2024 survey with the Bank of England
• Gartner, Over 40% of agentic AI projects will be canceled by end of 2027, 25 June 2025
• Themes from the Gartner Finance Symposium 2026, 16 June 2026
• Model Context Protocol, adoption and governance
• Microsoft WorkLab, How AI is redesigning work at BNY
• Gibson Dunn, EU AI Act Omnibus agreement, May 2026
• Elevate, SR 11-7 rescinded, on SR 26-2 and OCC Bulletin 2026-13
• Gradient Labs, back office AI platforms, vendor list and customers