Most of what we billed for was translation
A large share of platform consulting revenue has been paid for translation. Someone describes a requirement in a workshop, and a consultant turns it into configuration in one specific product: the right table, the right security model, the right approval step. The skill was real and so were the rates, which paid for the fact that every product hides what it can do behind its own vocabulary.
That is precisely the work generative tooling does best. Describe the outcome, get a plausible first configuration. The premium attached to knowing where a setting lives is falling. Anyone who has watched an agent build a working flow and a matching security role from a plain English paragraph knows this, and it is the shift behind what the admin job becomes once agents do the clicks.
That is a description of what is already happening rather than a forecast. Plenty of practitioners are still busy and the backlog has not vanished. What has changed is the argument you can make at renewal. Hours booked against build are harder to defend when the client can see how fast the first pass now is.
Deciding what should be built at all
The first thing holding its price is deciding what should be built. Requirements are not discovered by asking people what they want. You find them by watching what people actually do, finding the spreadsheet that shadows the system, and noticing that two teams describe one process with different nouns. A generated configuration answers the question you typed, and typed questions are usually wrong on the first pass.
Most of the value in a good consultant is the questions that stop the wrong thing being built. The new custom object that should have been a field. The approval chain that exists because a director was burned once and nobody has rechecked whether the risk still applies. None of that shows up on an invoice line, which is why juniors rarely learn to do it.
The structure that makes this work is a proper discovery conversation with the right people in the room and someone willing to write down what was agreed. If you want to get better at it deliberately, run a discovery workshop that surfaces real requirements instead of a session where everyone reads a wish list aloud.
Work where being wrong is expensive to check
The second thing holding its price is anything whose correctness is expensive to verify. Money, entitlement, identity, anything a regulator can ask you about later. Generated output is cheap and confident, and those two together are the problem. A plausible tax determination, a plausible commission split, a plausible access rule: each one looks finished and none announces that it is wrong.
Someone has to be accountable for the result being right, and accountability carries a price. The review and the sign-off hold value while the build around them gets cheaper. A client can generate a payroll calculation in an afternoon. Proving it pays the same net amount as the current system for every awkward case in the population takes a month, and that month gets paid to a named person.
If your current role is mostly build, move toward the verification end. Learn how the numbers get reconciled. Learn which controls the auditor tests and how. That knowledge does not get cheaper when configuration does, because the signature on the control still has to belong to a person.
Two systems disagreeing about what a customer is
The third thing holding its price lives at the boundaries. Single-product tooling sees one product, so it never sees that the CRM thinks a customer is an account with a billing address, the ERP thinks a customer is a sold-to party with a credit limit, and the service desk thinks a customer is whoever raised the ticket. A vendor agent will not surface that disagreement, because from inside one product there is none.
The hard problems live in those gaps. Which system owns the record, which one wins when they differ, and what happens to the downstream reference when two records are merged. Answering that requires holding two data models in your head at once and having opinions about both. The design questions behind what a shared customer record really needs have no product answer.
The same applies to the worker record, the asset record, anything that exists in two places with two lifecycles. Middleware moves the data. Deciding what the data means is a judgement call that survives every generation of tooling, because the disagreement sits between two business processes rather than two APIs.
Getting two departments to agree a definition
The fourth thing holding its price is political, and it is not a technical problem. Two departments have to agree what an active customer means, or which date counts as the hire date, or whose number goes in the board pack. No tool settles that. Someone sits in the room, hears both positions, and writes a definition both sides will defend a month later.
The work is slow and unglamorous and it is the reason projects land. A consultant who can run that conversation without either department feeling rolled over is worth more than one who can build faster, because the build was never the constraint. Arguments such as who gets to define revenue get settled by people with authority and patience.
Agents make this more valuable rather than less. An agent needs a definition before it can act on one, and a vague definition produces confidently wrong behaviour at volume. The organisations getting real value from agents are the ones that did the boring definitional work first.
Fixed bids change who gets paid
When delivery effort falls and becomes predictable, fixed-bid pricing spreads. It is already spreading at the edges of the market, because a client who believes the build is fast will not accept open-ended time and materials. Fixed bids move risk to the supplier. They reward firms that scope well and punish firms that staff heavily.
That is bad news for the traditional pyramid. A model that made its margin by putting billable juniors on configuration tasks under thin senior oversight has little to sell once those tasks are the cheap part. Small senior teams do better, because the scarce input is judgement, and judgement does not scale with headcount. Scoping accuracy becomes the commercial skill, which is why scoping a statement of work for a platform migration deserves the attention.
It is worse news for people whose experience is broad but shallow. Someone who has touched six products at surface level used to be useful, because surface familiarity across products was scarce. Generated configuration is very good at surface familiarity. Being the second-best option at six things has become a weak place to stand.
What to get good at over the next six months
Depth in a domain beats breadth across products. Knowing how revenue recognition actually works, or how a shift pattern turns into a payslip, or how a warranty claim gets assessed, travels across platforms and does not get generated away. Time spent understanding the business process pays better than time spent memorising screens, and the screens change anyway.
Being the person who can say no to a requirement and explain the reasoning is a durable position. It needs enough knowledge of the domain to be confident and enough history with the client to be trusted. Neither of those comes out of a model.
Grounding and context design is genuinely new expertise, and it sits closer to data architecture than to configuration. Deciding what an agent may read, how the source data is shaped, which records are authoritative and how stale is too stale is data work with a new consumer. Anyone who has cleaned up a configuration database before letting tooling read it recognises the work, and getting the CMDB right before agents read it is that discipline applied earlier.
Pick one process at your current client that you cannot yet explain end to end without opening the system. Learn it properly between now and spring, from the trigger through to the money, including what finance does with the output. Then write down the three questions you would ask before anyone configures anything in it. That page is what you will still be charging for in five years.
