An organisation decides it needs AI capability and immediately faces a question nobody prepared it for: who does the work. There are four realistic answers, they differ in cost by roughly a factor of ten, and the cheapest one on paper is frequently the most expensive one in practice. Almost nobody in this market publishes the comparison, so here it is with the numbers stated and their basis shown.
Key Takeaways
- Four routes: hire, traditional consultancy, offshore development, and embedded or forward-deployed engineering. Each is right for a different problem.
- A ten-day scoping exercise at international consultancy day rates can cost more than a month of embedded engineering, and produces a document rather than a system.
- Hiring is the cheapest steady-state option and the slowest and riskiest starting option, because the first hire has nobody to learn from.
- The line that decides the whole comparison is rarely on the quote: who maintains the thing afterwards.
- A build handed to an organisation with nobody to run it has a twelve-month lifespan and a rebuild at the end of it.
The four routes, priced
| Route | Published or benchmarked rate | Time to first working thing | What you own at the end |
|---|---|---|---|
| Hire an internal person | Market salary plus recruitment, onboarding and the risk of a wrong hire | Three to six months before output | Capability, if they stay |
| International consultancy | USD 430 to 1,100 per day, per United Nations published bands for 2025 | Weeks to a document | A strategy, an assessment, a roadmap |
| Offshore development | USD 6,000 to 14,000 per developer-month, the band published by large African talent marketplaces | Six to twelve weeks | Software, and a dependency |
| Embedded or forward-deployed engineering | From USD 4,500 to 6,500 per engineer-month at African-market economics | Weeks, shipping continuously | A running system, documentation and trained staff |
Those are honest bands rather than quotations, and the ranges are wide for a reason: scope, data quality and integration count vary more than rates do.
Where each one is genuinely the right answer
Hire, when the work is permanent and you already have someone to learn from
Hiring is the cheapest steady state by a distance. It is also the slowest start and carries a specific failure mode that this market has seen repeatedly: the first AI hire in an organisation with no existing capability has nobody to learn from, nobody able to evaluate their work, and nobody to notice when they are stuck. They frequently spend six months building something impressive that nobody asked for, and then leave for a role where their skills are supervised.
Hire when you already have a technical function that can absorb and direct the person, when the work is permanent rather than a project, and when you can wait a quarter for output. Do not make your first AI investment a job advert.
Consultancy, when the question is genuinely strategic
Consultancy earns its rate when the problem is a decision rather than a build: whether to enter a market, how to structure a function, what a regulator will expect. It stops earning its rate when what you actually needed was software, because the deliverable is a document and the document does not run.
The arithmetic is worth doing before commissioning. A ten-day scoping exercise at the top of the published UN band is USD 11,000 and produces a report. A month of embedded engineering at USD 4,500 to 6,500 produces a working system and a person on your team who now understands it. Both are legitimate purchases. They are not substitutes, and they are frequently confused.
Offshore development, when the specification is genuinely settled
Offshore works when you can write down what you want with enough precision that a team elsewhere can build it without asking you forty questions a week. That is a real condition and most AI projects fail it, because in AI work the specification emerges from contact with the data. You do not know what the system should do until you have seen what your records actually look like.
The other cost is timezone. A question that takes ten minutes to answer takes a day to round-trip, and AI projects generate a great many questions.
Embedded engineering, when the specification will emerge from the work
Forward-deployed engineering means the engineer works inside your operation, in your timezone, on your systems, and there is no handoff to an implementation team afterwards. The people who scope the work build the work and stay until it runs. Your constraints, meaning connectivity, legacy systems, data quality and budget, are treated as design inputs rather than as excuses.
This is the model Palantir invented and that the frontier AI labs adopted for their largest customers. Almost nobody offers it in East Africa, which is a market gap rather than a mystery: it requires people who can both build and sit in a room with a finance director, and those people are rare and mobile.
It is the right answer when the problem is real but the specification is not yet, which describes most first AI projects. It is the wrong answer when you know exactly what you want and simply need hands, in which case offshore is cheaper.
The three questions that move a quote more than the rate does
1. How clean is the data?
A build against clean, exportable records is a fraction of the same build against scanned PDFs and inconsistent spreadsheets. Data preparation is frequently the majority of a project and is routinely quoted as if it were a preliminary. Be suspicious of any proposal that jumps to the interface with no line for understanding and preparing your records. It has either not looked at your data or has decided to discover the problem after you have signed. Ask what happens to the price if the data turns out worse than expected, and get the answer in writing.
2. How many systems must it touch?
One system is a project. Four systems, two of which have no usable interface, is a different project wearing the same name. Integration count is the strongest predictor of overrun in this work, ahead of ambition and well ahead of technology choice.
3. Who maintains it afterwards?
This is the question that decides the comparison, and it is the one most often left to the end. A system handed to an organisation with nobody able to run it has a twelve-month lifespan and a rebuild at the end of it. That converts an apparently cheap build into the most expensive route on the table, because you pay twice and lose a year in between.
Whatever route you choose, name the maintainer before the work starts, not at handover. It changes the design, and it should.
A comparison worth running on your own numbers
Take one concrete project and price all four routes honestly, including the lines nobody quotes: your own staff’s time, the checking time on outputs, and the cost of maintaining it in year two. Most organisations that run this exercise discover two things.
The first is that the internal-time line is one of the two largest costs in every route, and it appears on no vendor quote. An organisation that budgets the licences and the fees but not the person has budgeted the cheap half, and that is the most common reason a well-funded AI programme stalls: the money was there and the hours were not.
The second is that the routes are not really competing. Consultancy answers a question. Offshore builds a settled specification. Hiring builds permanent capability. Embedded engineering finds out what should be built by building it. Most organisations need two of the four in sequence, and the expensive mistake is buying the wrong one first.
What we tell organisations that are not ready to build anything
Do the configuring properly first. Paid seats for the third of staff who need them, real training, a written policy, a tool register, and two honest thirty-day pilots. That costs a fraction of any build on this page and it tells you exactly what to build, or that you should not build anything yet.
A significant number of organisations that arrive wanting a custom system leave with a subscription, a policy and a much smaller invoice. They are better off for it, and so is the eventual build, because when it happens it is aimed at something real.
Sources
- United Nations published consultancy day-rate bands, 2025, used as an external benchmark.
- Published developer-month rates from large African talent marketplaces, USD 6,000 to 14,000.
- AVODA AI Consulting published rates for readiness assessment and forward-deployed engineering.
Rates checked 21 August 2026. Nothing here is a quotation and all bands move.