Hire in house or use an AI agency, and how to tell which one you are

The short answer, before the arguments

Hire in house when the AI work is continuous and central to how the business makes money. Use an agency when you need a defined set of systems built, running and handed over, and there is not enough ongoing change to keep a salaried specialist occupied. The test is not cost. It is whether the work has an end.

An employee is standing capacity. You pay for their time whether or not there is a queue in front of them, and in exchange you get somebody who learns the business, sits in the room when a decision is made, and is there the moment something breaks.

An agency is a bought capability. You pay for a set of systems and a period of support, and in exchange you get people who have built the same pattern elsewhere, at the cost of never knowing your business as well as your own staff do.

Most owners ask which is cheaper. That answer only holds for the first year. Ask instead how much genuinely new AI work you expect in the second year and the third, because that is the question the two models answer differently.

When hiring in house is the better call

There are situations where a retainer is the worse deal and no honest agency should pretend otherwise. The common thread is volume of change. If your systems are going to keep moving, somebody already inside the business will move them faster and more cheaply than any external arrangement can, because half of every external conversation is context that an employee already has.

The clearest case is when AI sits in the product rather than in the back office. A business whose customers interact with a model, whose pricing depends on one, or whose position against competitors rests on how well it works, should own that capability with a payroll number attached to it.

• The work never finishes. New processes, new data sources and new models arrive faster than anyone can write a scope for them.
• There is enough of it to fill a week, every week. If you can list a year of work and it keeps one person occupied, a salary is the cheaper unit of buying it.
• The capability is core. If a competitor could copy your automation and you would lose something real, that knowledge should not live outside the building.
• The domain takes months to learn. Unusual pricing rules, negotiated contract terms and processes that exist only as habit all favour somebody who is present for the conversations.
• Someone senior already understands the work well enough to hire for it and judge the output. An AI hire with no manager who can tell good work from confident work is an expensive experiment.

When an agency is the better call

The case for buying it runs the other way. The work has a shape, it needs to exist by a date, and once it exists somebody needs to notice when it breaks rather than keep inventing more of it.

Hiring carries a cost that rarely appears in the comparison. The weeks spent shortlisting, the manager hours lost to onboarding, and the months before anyone works at full speed all happen before the first workflow runs. A team that has built the pattern before starts from a working version of it, which is most of what you are paying the premium for.

There is also a skills problem that one hire cannot solve. A working system usually touches messaging platform rules, an automation platform, a CRM, a reporting layer and some model work. People who are genuinely good at all of those are rare, expensive, and generally not looking for a first AI role at a mid-sized company.

• You have a defined set of processes to automate and a date by which they should be running.
• You cannot yet write the job description, which is the honest state of most businesses at this point.
• The ongoing change is real but small: a few adjustments a month rather than a roadmap.
• You want the option to stop. A contract ends more cleanly than a role does, and everybody knows it.
• You need several specialisms at once for a few months and none of them at full time afterwards.

The criteria that actually separate them

Set the two beside each other on the things that change the answer, rather than on the things a proposal usually compares.

Criterion	In-house hire	Agency
What you are buying	Standing capacity, available whether or not there is work	A defined capability for a defined period
Time to first working system	Slower: recruit, onboard, learn the business, then build	Faster: the pattern already exists and gets fitted to you
Breadth of skills	One person's range, which is narrower than the work	Several specialisms, none of them full time
Where the knowledge sits	Inside the business, and it leaves when they do	With the vendor, unless handover is written into the contract
When the work dries up	The salary continues	The engagement ends
When something breaks at midnight	Depends on one person being reachable	Depends on the support terms you signed, so read them

What neither option fixes

Both answers assume the process is knowable. If nobody can describe how an order moves from enquiry to delivery, and three people describe it three different ways, then whoever you engage spends the first phase writing it down. You pay for that phase either as salary or as fees, and it is usually longer than anyone expects.

Neither helps if the choice is really a way of avoiding a decision. Hiring somebody to own AI without giving them authority to change how a department works produces a well-paid observer. Engaging an agency without naming an internal owner produces a system nobody inside the business can explain a year later, which is the same failure wearing a different hat.

There is a floor under both. A task that runs a handful of times a month rarely repays the cost of building and maintaining anything to do it, whoever does the building. Count the repetitions before you count the quotes.

A rule you can apply this week

List every piece of AI or automation work you can name for the next twelve months. Put an honest week estimate beside each one and add them up. Then mark which items you could still describe accurately if the person who suggested them left tomorrow.

If the total fills most of a year and keeps growing while you write it, hire. If it fills a quarter and then stops, buy it, and make handover a term of the contract rather than a favour you ask for later. If the list has three items and you are not confident any of them is the right one, do neither yet: spend a fortnight watching where the hours actually go, and the list will rewrite itself.

Twelve months of work, and the thing that outlasts whoever does it

The rule at the end is a twelve month list, so here is what the first two months against such a list look like from a supplier's side, published rather than estimated. The audit takes the first week. One named system is live inside a fortnight. The second and third land across month two. If your list holds four items for the year, that pattern finishes them and then there is not enough change left to keep a salaried person occupied, which is the article's own test answering itself without anybody arguing.

Neither answer removes the need for somebody inside the business standing at the stop. Spending, pricing, anything published in your name and any serious complaint wait for human approval, and a workflow that meets a case outside its scope hands it to a person with the record attached. An agency can build that escalation and cannot be it. Where nobody on your side will read what comes through it, the twelve month list is the wrong thing to be counting in the first place.

The case for buying rests on what still runs after handover, so it should be checkable rather than asserted. More than 500 calls a day at Big Texas Land Buyers with automatic CRM categorisation. Appointment booking largely hands-free at Center for Sight in New York. Three purpose built ERP systems live at RM Gulistan Engineers in Karachi. Ibrahim owns build and workflows at Wobble, and his stated goal is tools simple enough that a client's own team can be trained to run them without him, which is the agency case arguing quietly for its own ending.

Common questions

Is hiring in house cheaper than an AI agency?

Over one project, usually not, because the recruiting time, onboarding and ramp-up all land before anything works. Over several years of continuous work, usually yes, because a salary buys a fixed amount of capacity while fees scale with how much you ask for. The crossover depends on how much new work you actually generate each month.

What kind of person should I hire if I go in house?

Someone who can build automations and integrate systems, rather than someone who only prompts models. The scarce skill is connecting a CRM, a messaging platform and a reporting layer so that the whole thing keeps running unattended. Model work is the smaller part of most business AI, and the easier part to learn on the job.

Can I start with an agency and hire later?

Yes, and it is a common sequence. It works if the contract says you own the accounts, the workflows and the documentation from the start, and if the systems are built somewhere your future hire can read them. It fails when the work sits inside a vendor platform you cannot export, because then the hire inherits a black box.

How do I know if there is enough work to justify a hire?

Write the twelve-month list and estimate it in weeks. If it fills most of the year before you have finished writing, there is enough. If you struggle to fill a quarter, a salary will be paying for idle capacity, and the idle capacity tends to get filled with work that was never worth doing.

What should an agency contract include so I am not locked in?

Accounts and licences in your name, workflows and credentials on infrastructure you control, written documentation of what each system does, and a defined exit that hands over everything without a negotiation. Ask the plain question before you sign: if we stop next month, what do we keep?

Does an in-house hire make sense if we have never automated anything?

Rarely as the first move. Automating a process you have never examined means encoding guesses, and a new hire with no existing systems to learn from spends their first months doing discovery you could have bought faster. Get one or two processes working, see what the pattern of change looks like, then decide whether it needs a permanent owner.