How to Clear AI-Assisted Work: What Business Leaders Can Learn from Payments

Jason Gu, former Visa head of AI, describes “clearing” AI-assisted work: the payments approach to approving output, containing errors, and deciding who pays when AI gets something wrong.

Artificial intelligence (AI) can draft, extract, and summarize in seconds—but trusting the results hasn’t gotten cheaper.

Consider a recent anecdote. An organization adopted an AI agent to draft responses to technically complex customer support tickets, aiming to halve human handling time. But savings didn’t come to fruition—and handling time had shot up—because 10 percent of the AI’s responses required verification to avoid data loss for customers, and employees had to read and verify every line of generated text.

Fortunately, one industry’s precedent offers a solution to the AI-verification conundrum: payments.

Think about it. A store sells goods without verifying the customer because their credit or debit card was approved. Approval provides permission for the vendor to proceed—backed by rules about who bears the loss for incorrect approval. Checks, permissions, and consequences work together ensuring purchases go through without requiring manual review.

Jason Gu, a managing director in BRG’s AI & Decision Intelligence practice, led Visa’s development of AI platforms and products. Now, he advises clients on AI strategy, implementation, delivery, and governance.

To get started, how can AI-assisted workstreams adopt payments-style processes to reduce costs and improve accuracy?

Jason: AI-assisted work needs its own version of approval, which I call clearing. Cleared work has passed agreed checks and may be used for one stated purpose without another full review. Importantly, it isn’t certified correct or cleared for any other use and can be reopened.

The framework fits recurring, high-volume work that has a clear standard, affordable checks, and uses that can contain the consequences of a missed error. If a serious error could pass unnoticed and cause harm the process can’t contain, the work gets expert review, a narrower use, or no AI at all.

Predict what you can, contain what you can’t.

Your approach combines real-time risk scoring with clear liability frameworks, standardized dispute-resolution processes, and tokenization. What does each entail?

Jason: Imagine a private equity firm buying a business with 5,000 contracts. Its advisors use AI for a document index, key-term extraction, and the ultimate decision to buy.

The four elements you mentioned map to four crucial steps in clearing such work:

1. Agree to the rules (liability framework). Before work starts, the PE firm and advisor agree on:

• • the task
  • the use for each result
  • which checks clear the work
  • the quality standard
  • who investigates, fixes, and bears losses for each kind of failure.

The AI provider’s commitments—including keeping records, giving advance notice of changes, and cooperating when something goes wrong—must support these rules.

2. Test, then clear what qualifies (risk scoring). Test the process on fresh contracts that experts have already reviewed. Measure wrong answers, missed items, and important items wrongly routed toward clearance.

Test what gets through and what gets flagged. Work that meets the standard is cleared for its stated use with ongoing sampling to confirm the process keeps performing.

3. Escalate what doesn’t qualify. Business knowledge decides what must go to experts: in this case, top customers, sole-source suppliers, and high-consequence terms. AI can add work to the expert pile but never remove it. Missing pages, unreadable files, and unfamiliar contract types are held for review. Nothing is cleared by default.
4. Contain, record, and correct (tokenization and dispute resolution). Payment tokens limit where a credential works; cleared, AI-assisted work needs the same technical limit. For example, the index feeds document navigation, not the valuation model. Access and approval controls, rather than a label, enforce that limit.

Every cleared item keeps a record of its source, system version, checks run, and reviewer so disputes are settled against the agreed-upon rules. One serious miss is enough to act on: pause the affected work, notify the people relying on it, and recheck anything already released.

Many companies report that their AI pilot worked but the savings never materialized because of review and verification. Is that a technology or an economics problem?

Jason: From the outside both look the same: expert hours keep growing with volume. An unreliable tool forces extra review. A workflow without agreed standards makes people recheck work because “good enough” wasn’t defined and no one said who answers if the work is wrong.

AI can save effort in three ways: faster preparation, faster review, and clearing work. The first two make each review cheaper. Only clearing breaks the link between volume and expert hours, and that is where these processes can scale. Leaders should:

• Measure the cost of finished, acceptable work.
• Count every cost: software, testing, checks, expert review, corrections, insurance, and losses no one else covers.
• Compare the result with today’s process at the same quality and risk.
• If the full process isn’t better, narrow the task or stop.

What did payments stop trying to verify, what replaced it, and how does this apply to AI-assisted work?

Jason: Payments didn’t stop verifying; its checks keep getting stronger. It has never required proof before approving a purchase that said the purchase is legitimate and the merchant will deliver. No check could supply that proof in a second.

Instead, each technology—chips, tokens, online authentication—does a specific job. The approval lets the purchase proceed. Any later dispute follows separate rules that decide who bears the loss; for certain cases of fraud, specified security steps shift that responsibility.

AI-assisted work should have a rule for clearing. Checks create scale only when passing them grants permission to proceed, which is credible only when everyone knows what happens if it turns out to be wrong.

How do you decide which AI outputs need to be verified and which should just be priced as expected error? What cases must you verify, whatever it costs?

Jason: Price error only where you have evidence of how often it happens and a use that limits what it can cost. Verify everything else. Ask:

• What might this process miss? Test on real work, including omissions.
• What could a miss cost before it’s caught? Answer this from business knowledge and the rules governing the work.

Verify in situations where:

• a miss could cause severe or irreversible harm that the use can’t contain
• law or professional rules require review
• the process hasn’t been tested on that kind of item

Sampling has a hard limit. If one of 5,000 contracts contains the critical clause, a random sample of 100 has only a 2 percent chance of including it. Samples can’t certify that nothing serious was missed.

Who is liable when AI-generated work product is wrong?

Jason: Every AI-assisted workflow can fail in four places; the parties should agree in advance about who will investigate and fix the problem and how losses will be handled:

• Bad inputs. The information was wrong, incomplete, or withheld (e.g., a seller never uploads an amendment). This is usually the supplier’s responsibility unless the process was designed to catch the gap and didn’t.
• Failed controls. The process was inadequate, wasn’t followed, or wasn’t updated when things changed. Perfectly following a weak checklist still counts as a control failure. In payments, skipping required security steps can make you liable for certain fraud losses.
• Residual error. Something went wrong even though an appropriate process was followed. This is the cost of clearing work without reviewing everything. Someone must own it, or everyone rechecks everything and nothing scales.
• Misuse. The work was relied on beyond its cleared purpose. This is usually the responsibility of whoever extended the use.

The agreement should also say how shared losses will be handled. Two boundaries apply. First, an agreed error rate is a quality standard, not a liability waiver, and professional rules restrict how far advisors can limit their liability. Second, a bad outcome such as a customer leaving after closing doesn’t by itself prove the work was wrong.

The goal is less repeated work, not less responsibility.

Chargebacks work because card networks can impose rules on every participant. Who plays that role for AI-generated work product?

Jason: For the recurring services this framework targets, the client and service provider use a standard agreement for the task rather than negotiating afresh for each job.

Card networks work because every bank and merchant accepts one rulebook as a condition of participating, with federal consumer protections as a floor. AI-assisted work has no such network, but it doesn’t need one to get started. A service agreement can define:

• the quality standard and permitted uses
• what evidence settles a dispute, who decides, and how fast
• responsibilities for each of the four failure types listed above
• what the AI provider commits to, plus the technical setup that enforces it

The agreement must be honest about gaps. A fee refund doesn’t compensate for losing a major customer. The agreement should state what protection exists, whether that protection can be honored, and what risk the client still carries. Shared industry terms will follow as large clients standardize their agreements and insurers set conditions.

Which industries might this framework apply to?

Jason: Look for tasks, not whole industries. The right task is recurring work with a clear standard, affordable checks, and a use that can contain a missed error. The strongest candidates are in industries that already treat human work this way: they clear routine items, escalate exceptions, and audit by sampling.

By contrast, open-ended advice, one-off strategic decisions, and medical diagnosis are poor fits because the standard is contested and the harm is hard to contain.

Adoption will come from clients that want lower cost without lowering standards. They will ask: What did you clear, for what use, on what evidence; and who answers if it’s wrong?

Payments didn’t scale by trusting everyone but by agreeing when a transaction is good enough to approve—and what happens when it isn’t. AI-assisted work can scale the same way.