The agency types AI helps vs the agency types AI threatens
Most predictions about AI and agencies sort firms by how technical the work looks. That axis predicts nothing. A software shop and a transcription service both look technical from the outside, and their exposure could not be more different. The axis that actually predicts is simpler and less flattering: what does the client pay for, and can that thing be written down precisely enough to hand to a machine?
If the answer is a describable artifact produced mostly by junior labor, the category is exposed. If the answer is accountability, a license, a body in a building, or a relationship a person has to keep alive, the category is not exposed on the revenue side, and the pressure lands on its costs instead. That is the whole argument. The chart below sorts categories on that one question.
Start at the top, and start honestly. Transcription is the clearest case in the entire agency economy. The client buys a text file that matches an audio file. The artifact is fully describable, the quality is measurable against an obvious standard, and machine output now lands close enough that the human role has moved from producing to reviewing. The same shape holds for translation and localization, presentation design, copywriting, production-heavy content marketing, and data entry and back office work. In every one of them the cost of goods is junior labor turning an input into a deliverable that can be specified in a sentence.
The mechanism that compresses price is not quality. It is visibility of the cost floor. A buyer who has personally watched a model produce a decent draft now carries a rough number in their head for what raw output costs. Every quote gets measured against that number, including quotes for work the model could not actually finish. This is why good firms in these categories report the same strange year: their output is better than it has ever been, their clients are happier than they have ever been, and their rate card is under pressure anyway.
Displacement here is real and it is worth naming plainly. The people most affected are not agency owners. They are transcriptionists, junior translators, entry-level copywriters, deck builders, and offshore back-office teams, and those roles are thinning. Nothing about this article should read as a shrug at that. The firms in these categories that hold their pricing are the ones that sell something the artifact never was: certified accuracy, admissible records, regulated medical or legal translation, sworn attestation, a named human who signs off and carries the consequence of being wrong. That is a smaller business than the one they had, and it is a real one.
Outsourced bookkeeping sits at the hinge of the chart and shows how fine the line is. Transaction categorization and bank matching, which used to be most of the hours, are now largely automated, so the hours-per-client number has been falling for a while. But the client is not buying categorized transactions. They are buying books that survive an audit, a clean close each month, and a person who notices that payroll ran twice. The first half of that job is gone and the second half is not, which is why the honest description of this category is not collapse but compression: the same work, fewer hours, and firms that either take on more clients per bookkeeper or watch revenue drift down.
Now the other end of the chart, and be precise about why it looks the way it does. A home care agency is not selling a care plan document. It is selling a background-checked, bonded, supervised person who arrives at your mother's apartment on Tuesday morning and is still there when she needs help standing up. AI absorbs a great deal of what that agency does internally: caregiver matching, shift scheduling, intake paperwork, visit notes, state compliance documentation. It absorbs none of the thing the client is buying. Cost per hour of overhead falls, the billable hour stays human, and the gap between them is margin.
That same shape repeats across home health, security guard agencies, freight brokerages, temp staffing, insurance agencies, executive search, private investigation, and property management. The common thread is that each of these firms carries something on behalf of the client. A payroll and its tax risk. A cargo liability claim. A state license that can be revoked. A duty of care to a vulnerable person. A fiduciary relationship with a carrier. Software produces output; it does not carry risk, and it cannot be the entity that gets sued or fined. Someone still has to sign, and the signature is a large part of the price.
Exposure is not about how technical the work is. It is about whether the client can describe what they are buying.
The middle of the chart is where most of the bad analysis happens. SEO, paid media, MSPs, web development, and ecommerce agencies get called either doomed or safe, and neither is right, because the category is not the unit of exposure. The billing model is. Production cost inside these firms has collapsed, and the value of judgment, sequencing, and accountability has gone up. A shop that bills hours has just watched its inventory lose value. A shop that sells a result at a fixed price has just watched its cost of delivery fall while the price holds. Two firms in the same category, opposite outcomes, and the difference was decided by a contract clause years before any of this.
MSPs are the sharpest illustration. A lot of MSP labor is triage, documentation, patch summaries, and first-line ticket response, and that work is genuinely absorbable. But the client is not buying tickets. The client is buying someone whose phone rings at two in the morning when the server is down and who is contractually on the hook for getting it back. Cost per endpoint falls; the monthly contract does not, because the contract was never priced on effort. The same logic protects the ecommerce agency that owns a revenue number and squeezes the one that bills for store maintenance.
Two things cut the other way, and both deserve to be stated without inflation. The first is that AI has created a real new category. AI and automation agencies are among the fastest-growing agency types in the country, and demand for them is coming from firms that know they need to change something and have no internal capacity to do it. Plenty of that market is repackaged consulting that will not survive contact with a second year. The ones that look durable look structurally like MSPs: they build something, they maintain it, they charge monthly, and they are accountable when it breaks.
The second is that search behavior is shifting toward AI answers, which changes what SEO agencies are actually selling. Optimizing to be quoted inside an answer is not the same job as optimizing to be clicked from a list of links, and the measurement, the deliverable, and the reporting all have to change with it. Some firms have already rebuilt around this. Others are selling rank reports for a page fewer people look at. It is too early to say where the equilibrium lands, and anyone telling you they know is selling something. What is observable today is that the firms treating it as a real change in the job are the ones still raising prices.
Which brings up the uncomfortable part of the word helped. An agency in an insulated category usually gets there by needing fewer junior people. The home care agency that ran six schedulers may run three. The search firm that employed a team of researchers employs two and better tools. Revenue per head goes up, margin goes up, and the owner correctly describes this as being helped. The person who would have been hired into the fourth scheduler seat describes it differently, and they are also correct. Being insulated on price is not the same as being insulated on headcount, and it changes hiring in a way the industry has barely started to talk about. The traditional agency pyramid depended on cheap junior labor doing describable work, which is exactly the work that got absorbed. Nobody has a good answer yet for where the next generation of senior people gets trained.
If you run an agency, the useful test is one question asked honestly. If a client wrote your deliverable down in a paragraph, could they get most of it without you? If yes, your pricing is exposed no matter how good your work is, and the response is not to argue about quality but to move toward whatever you can carry that a model cannot. If no, then name the reason precisely, because that reason, the license, the liability, the person who shows up, the signature, is now the entire business.