Brussels reached political agreement on the artificial intelligence regulation four days ago, after a negotiation that ran through the weekend. The largest consultancies are now reporting generative AI bookings in the hundreds of millions, and every outsourcing renewal conversation this quarter has arrived with a slide about it. What is actually changing is not the technology in the delivery centre. It is the shape of the team, and almost nobody is designing that deliberately.
The interesting question is not which tasks the machine takes. It is where you put the human, and what you hold them accountable for
Every hybrid design answers that question, usually by accident. The answer determines the cost, the quality and whether the arrangement survives its first serious error.
Three positions, and they are not interchangeable
In the loop. A person reviews every output before it takes effect. Maximum assurance, minimum leverage, and the default that quietly consumes the entire business case. If a human checks every invoice the model coded, you have added a step rather than removed one. On the loop. The system processes and acts; the human handles exceptions and samples the rest. This is the only configuration that produces real leverage, and it requires you to accept that some errors will pass unreviewed — which is a decision about tolerance, not a failure of design. Over the loop. The human owns the process rather than the transactions: monitors aggregate quality, investigates drift, adjusts thresholds, redesigns when the pattern changes. The common design error is inverted staffing. Experienced people end up in the loop reviewing routine output, because that feels safe, while the over-the-loop role — the one requiring the most judgement — is left to whoever has capacity. Put your best people over the loop and instrument the loop itself.
Partition the work three ways, not two
The framing of "human versus AI" causes more bad architecture than any other idea in this area, because it omits the component that should be doing most of the work. Give the model comprehension: reading documents, extracting fields, classifying, drafting, summarising, and flagging what looks unusual. Give the deterministic system everything with a right answer: arithmetic, matching, tolerance checks, tax treatment, posting rules, approval thresholds. Give the human judgement, exceptions, relationships and accountability. Almost every hybrid that fails in production failed because a language model was allowed to do arithmetic or apply a rule that should have lived in code. Models are excellent at understanding a messy document and unsuitable as a calculator.
| Component | Work assigned in the article |
|---|---|
| Model | Read, extract, classify, draft and flag unusual material. |
| Deterministic system | Apply arithmetic, matching, tolerances, posting rules and approval thresholds. |
| Human | Own judgement, exceptions, relationships and accountability. |
Qualitative summary of this article's source text, not a measured outcome or performance estimate.
Confidence routing is the mechanism, and the threshold is an economic decision
Score every item. Above the threshold it processes automatically; below it, a person sees it. The two numbers that then run your operation are the straight-through rate and the cost per exception. Set the threshold by arithmetic rather than instinct. What does an error cost — in rework, in interest, in a customer relationship, in a regulatory finding? What does a review cost? Where those curves cross is your threshold. Teams that pick a round number and defend it are usually running far too conservative, paying for review on a population where errors are cheap and reversible.
Accountability does not decompose
Whatever the split, one named person owns each output. No auditor, regulator or customer has ever accepted a distributed answer, and with the European text now politically agreed, documented human oversight moves from good practice toward obligation for higher-risk uses. Write the accountability map alongside the process design: for each output type, who is answerable, what they are expected to have checked, and what evidence exists that they did. If that document is hard to write, the design is wrong.
What it does to the provider relationship
The team you are buying changes shape. Instead of thirty processors in a delivery centre you are buying a platform plus eight senior people, which alters the risk profile in ways the commercial discussion usually misses. Concentration rises: far fewer individuals understand your process, and losing two of them matters more than losing ten used to. Knowledge transfer becomes harder, because the knowledge now lives partly in configuration rather than in people. And exit becomes materially harder, because the asset that makes the process work is no longer a runbook. So change what you ask for on exit. You need the configuration, the rule and prompt set, the confidence thresholds and their history, the exception taxonomy, and any labelled data produced from your transactions. A provider unwilling to commit to that is telling you something useful about the next renewal.
The exception taxonomy is the asset nobody writes down
The complete list of ways your process goes wrong, with frequency and cost attached, is the single most valuable input to a hybrid design and it almost never exists. Build it before you buy anything: three weeks of logging what actually gets escalated and why will tell you more about where automation pays than any vendor assessment. It is also the artefact that survives a change of provider, which is reason enough to own it.
Practical Guidance for AI-Human BPO Design Consultation
- Decide in, on or over the loop for each output type explicitly.
- Put experienced people over the loop, not in it.
- Keep arithmetic and rules in deterministic code, never in the model.
- Set confidence thresholds from error cost, not from a round number.
- Run the operation on straight-through rate and cost per exception.
- Write the accountability map before go-live.
- Build the exception taxonomy and keep ownership of it.
- Extend exit rights to configuration and labelled data.
The Regional Angle
Three things reshape this design for organisations operating here, and the first is structural. Most regional groups of any size do not outsource their back office to a third party; they run a captive shared services function, in Dubai, Riyadh, Cairo or India, staffed by their own people. That changes the hybrid problem in both directions. It is easier, because you own the data, the process and the systems, and there is no commercial negotiation about who captures the gain. It is harder, because a captive has no platform to bring, no product roadmap and usually no capital budget for one — so the sensible pattern is to buy the tooling and keep the process, rather than outsourcing wholesale to acquire capability. It also requires changing how the captive is measured. A shared services centre judged on cost per full-time equivalent will resist every change described here, because success now means fewer people doing harder work. Move the scorecard to straight-through rate, cost per exception and cycle time before you ask for a redesign. The second is the dominant source of exceptions in this region, and global process designs never account for it. Identity and entity matching here fails constantly on names: transliteration differs between the passport, the trade licence, the bank mandate and the invoice; Arabic and English forms of the same company name coexist; the legal entity on the licence differs from the trading name on the purchase order; and individuals appear with two, three or four spellings across your own systems. That single category drives a large share of matching failures in regional payables and receivables, and no confidence threshold will fix it, because the model is not wrong — the data genuinely does not match. Build the alias and transliteration table as a maintained master data asset with an owner, seeded from your existing exception history, before designing any routing. Without it your straight-through rate has a ceiling that has nothing to do with the technology you bought. The third is legal rather than operational. In many regional entities the people who may bind the company are a small named set, defined by the trade licence, the memorandum or a specific power of attorney, and that authority is not freely delegable. An on-the-loop design that assumes any qualified reviewer can approve an outgoing document may be operationally sensible and legally void for anything requiring an authorised signature — bank instructions, statutory filings, certain contracts and customs documentation among them. Map which outputs carry that requirement and design those paths as in-the-loop by obligation, with the signatory's review evidenced. Everything else can be routed on economics. Discovering this distinction during an audit is considerably more expensive than spending an afternoon on it with your corporate services adviser.
The objection worth taking seriously
The strongest objection is that this is an elaborate design exercise for something that will be a product within eighteen months. The providers are building exactly these capabilities into their platforms, the software vendors are embedding them into the applications you already run, and a company that spends two quarters designing a bespoke hybrid operating model will find in 2025 that it could have bought the same thing configured. On that reading the only decision that matters is choosing a competent provider and letting them bring their tooling, and everything above is consulting that charges by the workshop. Much of that is right. Confidence routing, exception queues and document comprehension will be commodity features, probably sooner than eighteen months, and building any of it yourself would be a poor use of capital. But three things in this article are not purchasable, and they are the three that determine the outcome. The exception taxonomy describes your business, not the category, and no provider can hand it to you. The accountability map is a governance decision that cannot be delegated to a supplier, and the agreed European text is about to make that explicit. And the decision about where the human sits is where the entire economics of the arrangement is settled — a provider with a strong incentive to keep bodies in the loop will quite reasonably propose keeping bodies in the loop. Buy the platform. Own the design, the taxonomy and the accountability, because those are what let you change the platform later.
Common Questions
What straight-through rate should we expect?
It varies enormously by process and data quality, which is why the exception taxonomy comes first. Any number quoted before you have measured your own exceptions is a sales figure.
Should we keep a human reviewing everything at the start?
For a defined stabilisation period, yes — but with an agreed date and criteria for moving to sampling. Temporary full review that has no exit becomes the permanent operating model.
Does this reduce headcount?
It changes the mix more reliably than it reduces the total: fewer processors, more exception handlers and process owners, and a higher average skill level and salary.
What should we expect over the next twelve months?
Expect the European rules to be finalised in text over the coming months and to phase in over roughly two years, with the human oversight documentation requirements shaping supplier contracts well before they bite legally. Expect providers to arrive at the next renewal round with outcome-based AI proposals, and expect the first versions to be better for them than for you. Expect straight-through rate to become a standard line in service reviews during the year. And expect a visible wave of disappointing hybrid deployments around this time next year, almost all of which will have failed for the same reason: the human was left in the loop, and the arithmetic never worked.
AI-Human BPO Design Consultation — we build your exception taxonomy, set confidence thresholds from real error costs, and design where the human sits so the business case survives the first year.