For a small business that wants its first two AI agents live and running within a year, the twelve-month cost in India is roughly the same for a freelancer and an agency, about fourteen to fifteen lakh rupees in the worked example below, and about half as much again for an in-house engineer, who then costs that every year. The difference is not the price; it is what you own at the end, who carries the risk when the person leaves or the model changes, and how many hours of your own week the option consumes. Hire in-house when agents are becoming the product and you will have five or more by year two. Use a freelancer for a small internal tool you can scope and manage yourself. Use an agency for anything customer-facing, where a fixed fee, a written handover and someone to call matter more than the last rupee. Our prices are on this page as one agency’s example, marked as proposals where they are not yet confirmed.
Every option on this page can build a good agent; we have seen all three do it and all three fail. The failures are rarely about skill. They are about a hire who leaves at month seven with the only copy of the rulebook in their head, a freelancer who is unavailable the week the model is deprecated, and an agency whose handover is a PDF. So the comparison is in money, time and risk, with the arithmetic printed so you can swap in your own numbers.
| In-house engineer | Freelancer | Agency (our prices as the example) | |
|---|---|---|---|
| Time to the first agent live | 3–4 months: 6–8 weeks to hire, then a ramp on your systems | 4–8 weeks, if available when you are | 2–3 weeks from the scope call, as a fixed-price pilot |
| Year-one cost, India, two agents live (worked below) | About ₹22–23 lakh, then the same every year | About ₹14–15 lakh, assuming the hours hold | About ₹14 lakh including nine months of the Care retainer |
| How the price moves | Fixed salary whether or not anything ships | By the hour; scope creep is your bill | Fixed fee against a brief; changes logged and costed |
| Your hours a week | 3–5: managing one person, deciding priorities, being the product owner | 3–4: writing scope, reviewing, chasing | About 1, after the scope call |
| What you own at the end | Everything, if it was documented; often it was not | Whatever the contract says; the repository, usually | Repository, prompts, accounts and runbooks, IP assigned in the contract |
| When the person leaves | The agent stops improving; a replacement takes a quarter to learn it | You find another, who starts from the handover you received | Someone else at the agency reads the runbook; the handover was written for this |
| When the model is deprecated | Handled, if they are still there and not on something else | A new engagement, at their availability | Inside the retainer, or a scoped change without one |
| Guardrails, tests, audit log | As good as the one person’s habits | Varies widely; specify it in the scope or it will not be there | Standard on every build, published on the case studies |
| Best at | Many agents, continuous change, agents as the product | One well-scoped internal tool | Customer-facing agents with a measured result and a handover |
When agents are becoming how the business works, not a project: five or more workflows, rules that change weekly, integrations into systems only your team understands. The salary is justified by the volume of change, not by the first build. Hire after the first two agents exist, so the person inherits a working pattern and a log rather than a blank page.
When the agent is internal, small and well understood: a morning summary, a document check, a draft for a person to send. A good freelancer at a fair hourly rate is the cheapest way to get that built, provided you write the scope, hold the test set and insist on the handover in the contract.
When the agent talks to customers or writes to a system of record, when you want a fixed price against a brief, and when the thing you are buying is the handover, the guardrails and the thirty days after go-live rather than hours. The retainer after that is optional, and the honest version says so.
Assume a twenty-person business, a clinic or a CA firm, that wants a customer-facing agent (bookings or document collection) and one internal agent (the morning summary or a reconciliation check) live and maintained by the end of the year. Model usage and hosting are the same on every row, in the low thousands of rupees a month at this volume, so they are left out of the totals. Every figure is in the table; change the ones that are wrong for you.
| Line | In-house engineer | Freelancer | Agency (ours as the example) |
|---|---|---|---|
| Getting started | Hiring: 6–8 weeks and a fee or referral bonus, about ₹1.4 lakh at 8% of CTC | Finding and vetting: a week or two of your time | A free 30-minute scope call |
| The person or the fee | Mid-level AI engineer at about ₹18 lakh CTC (reported range ₹15–30 lakh); with provident fund and insurance about ₹19.8 lakh | About ₹3,000 an hour (reported range ₹2,000–4,000 in India for mid-level AI work) | Pilot ₹3,50,000 fixed; the second workflow from ₹6,50,000–₹35,00,000 (the low end is used below) |
| First agent live | Month 3–4, after hiring and ramp | 120 hours: about ₹3.6 lakh | ₹3,50,000, 2–3 weeks |
| Second agent live | Month 6–8 | 300 hours: about ₹9 lakh | About ₹6.5 lakh at the low end of the one-workflow band |
| Keeping both running | Included in the salary, when they are not on the next build | 8 hours a month for 9 months: about ₹2.2 lakh | Care retainer at ₹45,000 a month for 9 months: about ₹4 lakh; optional, 30 days of support are included anyway |
| Tools, laptop, accounts | About ₹1.5 lakh | Theirs | Ours, except your own cloud and model accounts, which you keep |
| Your hours over the year | About 200, as the engineer’s manager and product owner | About 150: scope, review, chasing | About 50, mostly in the first month |
| Year-one total (excluding your hours and running costs) | About ₹22.7 lakh, recurring | About ₹14.8 lakh | About ₹14 lakh |
| Year-two total, same two agents, no new builds | About ₹21 lakh | About ₹3 lakh of maintenance hours, if available | About ₹5.4 lakh on Care, or nothing if your team runs it |
The in-house row looks worst in year one and best in year three, and that is the honest shape of it: a salary is a terrible way to buy two agents and a good way to buy twenty. The question is whether the business will have twenty. Most twenty-person businesses will have three or four, with rules that change monthly rather than weekly, and a retainer or an occasional scoped change covers that for less than a salary and without the hiring risk.
The freelancer and agency rows look alike in money, which surprises people. The difference is in the columns that are not money: who holds the test set, whether the handover is written for a stranger, who answers when the model vendor sends a deprecation notice, and how many of your own hours the year consumes. A freelancer who provides all of that is an agency of one and should be priced like one; many excellent ones are.
The row everyone forgets is their own time. Two hundred hours managing an engineer is five full working weeks of an owner’s year. If the owner’s time is worth anything close to the engineer’s, the in-house total is nearer thirty lakh.
Buy the first one or two agents as fixed-price builds with a written handover and a test set. Run them for a few months on the retainer, or on your own, and watch the log: how often rules change, how many escalations, how many new workflows people ask for. If by the fourth agent the answers are "weekly, many, lots", hire, and hand the new engineer a working pattern, three logs and a runbook instead of a job description. If the answers are "monthly, few, two more", keep buying builds and keep the salary.
We write every handover for exactly that hire: the repository is yours, the prompts and rulebooks are documents, the test set runs in your CI, and the runbook assumes the reader has never met us.
The shape is the same abroad; the salary row moves most. Reported 2026 figures put a US AI engineer at roughly $140,000–$150,000 a year before the 25–30% employer overhead, and freelance AI developers at roughly $95–$235 an hour depending on seniority, against $25–$45 for comparable work billed from India. An agency working from India at fixed fees sits between the two, which is the arithmetic behind our USD prices: the pilot at $4,500 and one workflow in production from $8,000–$42,000. UK salaries sit below US ones and above Indian ones; use the vendor and aggregator reports for the current spread, and treat all of them as ranges from parties with an interest.
Answer them in order. Most people are done by the third.
Will the business have five or more agents, with rules that change weekly, within eighteen months?
Yes: Plan to hire, after the first two are live. Continue.
No: Do not hire for this. Continue.
Does the first agent talk to customers or write to a system of record?
Yes: Agency, or a freelancer who can show you a test set and a handover from a previous build.
No: A freelancer with a tight scope is fine.
Can you write the twelve-field scope and hold a builder to it yourself?
Yes: Freelancer is viable. Continue.
No: Agency; the scope is part of what you are paying for.
Is a fixed price against a brief worth more to you than the lowest hourly rate?
Yes: Agency.
No: Freelancer, with a cap in the contract.
Is there someone to manage the person weekly for a year?
Yes: Any of the three; pick on the rows above.
No: Not in-house, and not a freelancer without a very clear scope.
Only if the engineer is busy with agents for three years. Three years of a mid-level salary in India is fifty to sixty lakh with increments; that buys a lot of fixed-price builds and retainer months. If the business will genuinely have a dozen agents and continuous change, hiring wins, and the page says so. If it will have four, it does not.
Yes, and a good one will do it well. Ask for three things in writing: the tool list with permissions, the test set they will run before each change, and the handover document from a previous client with the client’s name removed. If all three arrive, the price difference between us is small and you should pick on availability and fit.
With us: the repository, the prompts and rulebooks as documents, your own cloud and model accounts, the runbook and the test set, with IP assigned in the contract and thirty days of support after go-live. Ask any agency the same question and read the contract rather than the website; the handover is the thing you are buying.
In a business with many agents: the weekly reviews, rule changes, model upgrades, new integrations and the next build, which is a full job. In a business with two agents: not much that needs that salary, which is why the page suggests hiring at the fourth agent rather than the first.
That is the pattern we recommend and build for. The handover assumes a future hire who has never met us: documented repository, rulebooks in sheets, a test set in CI, a runbook. A new engineer inherits a working pattern and three logs instead of a blank page, and the retainer ends with a month’s notice.
Aggregators report a mid-level AI engineer at roughly ₹15–30 lakh a year in 2026, more at global capability centres and product companies; add about ten percent for provident fund and insurance, a hiring fee or referral bonus, and six to eight weeks to fill the seat. Senior and generative-AI specialists are reported well above that. Treat the figures as ranges from parties with an interest.
Hire when agents are becoming how the business works and you can name five workflows for the next eighteen months; the salary is justified by continuous change, not by the first build. Use an agency when the agent is customer-facing, you want a fixed fee and a written handover, and nobody in the business will manage a developer weekly. The pattern that usually works is agency first and a hire at the fourth agent.
It depends mostly on where they are. US agencies are reported at $125–$240 an hour, which puts a single production agent at $60,000–$200,000; UK day rates are reported around £700 for contractors and from £800 for consultancies. Working from India, our fixed fees are $4,500 for the pilot and $8,000–$42,000 for one workflow in production, published on the pricing page and proposals until confirmed on a call.
Free, no deck. You leave with an approach, a timeline and a price range, or a reason to buy something off the shelf instead.