Contract, Permanent or Fractional: Choosing Your AI Engineering Employment Model in India and the UK

Three ways to sell the same skills

Almost every conversation about an AI engineering career is about where to work — which lab, which start-up, which global capability centre. Far less goes to the question that often matters more over five years: how you are engaged. The same skills can be sold into the same market in three structurally different ways.

Permanent employment is a contract with one organisation: salary, benefits, notice periods both ways and, usually, exclusivity. Independent contracting is selling your time at a day or hourly rate, through your own company or an intermediary, for a defined engagement with an end date, no benefits and no guarantee of what follows. Fractional work is selling a slice of your week — two days here, one there — to several clients at once, usually at a senior or advisory level, with the portfolio itself acting as your diversification.

In AI engineering the choice has sharpened rather than blurred. Demand is concentrated in a few specialisms — retrieval that survives real query distributions, evaluation infrastructure, inference cost work, agent orchestration, fine-tuning — and much of it arrives as short, well-specified engagements rather than permanent headcount. A bank in Mumbai or an insurer in Leeds needing an evaluation harness does not necessarily need an evaluation engineer for life. That is why so many good engineers now hold a contract offer beside a permanent one with no way to compare them.

This guide judges the three on four axes: total realised income (not headline rate), income stability (what happens in a bad quarter), learning and depth (what you get better at), and optionality (how easily you can change your mind).

Be clear before we go further: this is general career and commercial guidance, not tax or legal advice. The mechanics below are stable enough to plan around; the rates, thresholds and rules attached to them change, sometimes annually. Where a number matters to your decision, check the current position with HMRC, the Indian income-tax and GST authorities, or a qualified accountant.

Two things this guide is deliberately not. It will not tell you what to charge — that is our guide to freelance AI engineer rates and positioning — nor how to negotiate an offer, which is our guide to salary negotiation in a two-tier market. Read those for the numbers; read this for the structure they sit inside.

The comparison, honestly

Here are the three side by side. Treat the "typical" column as a central tendency in Indian and UK AI engineering as of August 2026, not a rule.

The point most people miss is a framing error, not a factual one. These three models are not a ladder. There is a persistent story that you start permanent, graduate to contracting, then ascend to fractional, as though each step were a promotion. Believing it makes people move for the wrong reasons.

Movement between them is normal and reversible in both directions. Engineers go independent for two years, learn five production stacks, then take a permanent staff role to own a system for longer than a quarter. Others go fractional after a decade employed, find the context-switching exhausting and go back. Neither is a demotion, and the hiring managers in Bengaluru and London who matter do not read them as one. The only pattern that reads badly is churn with nothing finished — a problem of outcomes, not structure.

The maths nobody does properly

The commonest error here is comparing a day rate to a salary by multiplying. A recruiter offers £550 a day; you multiply by roughly 260 working days, get about £143,000, compare it to your £78,000 permanent offer and conclude that contracting pays nearly twice as much. It does not. Two things break the multiplication.

The first is the pile of costs an employee never sees because an employer absorbs them: employer-side social contributions, paid leave, public holidays, sick cover, pension or provident-fund contributions, health insurance, hardware, software and cloud spend, professional indemnity insurance, accountancy fees and training. Each becomes a line item funded from the same day rate.

The second, and by a wide margin the largest, is utilisation. A day rate is not income; a day rate multiplied by billable days is income. And 100 per cent utilisation does not exist, for anyone, in any market. You do not bill the week you take off, the public holidays, the days you spent ill, the afternoon on your filings, the calls with a prospect who went quiet, or the six weeks between one contract ending and the next starting. A well-established independent might bill 200 days; a good first year is often closer to 140 — a 30 per cent swing on your whole income.

Here is the arithmetic done properly for a UK contractor. Every figure is illustrative, chosen to show the method — substitute your own, and see our 2026 AI engineer pay benchmarks for market numbers.

The illustrative permanent offer was £78,000 plus pension, leave and cover. Against £86,700 the contract is still ahead — by roughly 11 per cent, not 83 per cent, and that 11 per cent is your payment for carrying the risk of the gap. Drop utilisation to 140 days and the same rate yields about £64,700 comparable, materially worse than the permanent offer.

The Indian version has the same shape and different line items.

That last row matters for anyone new to invoicing in India: GST you collect is not income. It sits in your account looking exactly like income and it is not.

Now the counterweight, because none of this argues against contracting. Contract rates in genuinely scarce AI specialisms carry a real premium over permanent equivalents, because the supply of people who have shipped a working evaluation suite or a cost-controlled inference layer is thin — a scarcity visible in the 2026 talent-shortage and salary reporting. Short engagements also compound learning fast: four production RAG systems in two years teaches you more about what breaks in retrieval than one does in four. The honest answer is neither "contracting pays more" nor "permanent pays more" but it depends on your utilisation and your risk appetite — both measurable about yourself rather than guessable.

# break_even.py — the only spreadsheet you actually need.
# Currency-agnostic: put everything in one currency and stay consistent.

def break_even_day_rate(target_pre_tax_income, annual_business_costs,
                        billable_days):
    """Minimum day rate that clears your costs and income target."""
    if billable_days <= 0:
        raise ValueError("billable_days must be positive")
    return (target_pre_tax_income + annual_business_costs) / billable_days


def billable_days(working_days=260, leave_days=25, public_holidays=10,
                  sick_days=5, unbilled_business_days=40):
    """Everything that is not billable, stated explicitly."""
    return (working_days - leave_days - public_holidays
            - sick_days - unbilled_business_days)


if __name__ == "__main__":
    days = billable_days()                            # -> 180
    costs = 5000 + 2500 + 1500 + 600 + 1200 + 1500    # -> 12300
    print("billable days:", days)
    print("break-even rate:", round(break_even_day_rate(78000, costs, days)))
    # Sensitivity: what a bad year does to the same income target.
    for d in (140, 160, 180, 200):
        print(d, round(break_even_day_rate(78000, costs, d)))

The UK mechanics: IR35, umbrella and limited company

The UK tax system asks a question with no crisp technical answer: does this working arrangement look like employment, once you set aside the company structure in the middle of it?

That is the whole conceptual content of the off-payroll working rules, and the tests are behavioural rather than documentary. Control: does the client direct how, when and where you work, or do you? Substitution: do you have a genuine, exercisable right to send a suitably qualified substitute, or does the client require you personally? Mutuality of obligation: is there an expectation that work keeps being offered and accepted, or does the relationship end cleanly when the deliverable does? Contract wording matters less than how the engagement operates. A substitution clause nobody would honour is not a substitution right.

Two structural points matter. For medium and large private-sector clients, responsibility for the status determination sits with the client, not with you — which is why many large UK employers issue blanket inside-IR35 determinations rather than assess each engagement, and some will not engage limited-company contractors at all. And a small-company exemption exists, but be careful what it means: where the client is small, responsibility for assessing status stays with your own company instead of passing to the client. That is a change in who decides, not an automatic outside-IR35 answer — the rules still apply and the engagement can still be inside. It is nonetheless why the same engineer can end up assessed outside IR35 with a twenty-person start-up in Bristol and inside it at a Canary Wharf bank doing similar work. HMRC publishes off-payroll working guidance and a status-checking tool. Determinations are fact-specific and the rules are revised; treat any summary, this one included, as orientation rather than an answer.

In practice you will be routed down one of three paths. Your own limited company gives the most control over how you draw income, the most admin, and the most exposure if a determination goes against you — sensible for genuinely outside-IR35 work and for anyone with several clients. An umbrella company employs and payrolls you for the engagement: take-home is lower because employment costs come out of the assignment rate, but you get statutory rights, holiday pay and near-zero admin. Agency PAYE is similar, with the agency payrolling you directly.

Do not treat inside-IR35 as automatic disqualification. For twelve weeks of evaluation work at a Manchester health-tech, the umbrella route is often right: a limited company's overhead is real, the engagement is too short to amortise it, and the rate premium may still leave you ahead. Choose the structure to fit the engagement, not the reverse — and check who you are being routed through, because umbrella quality varies enormously and it is your payslip.

There is an AI-specific wrinkle. Highly specified, deliverable-based work — build this evaluation harness, deliver this fine-tuning pipeline — with real substitution rights and several concurrent clients tends to look more like a business-to-business supply than a long embedded engagement on a client's core roadmap. The engineer who spends eighteen months as de facto lead of a client's platform team, in their standups and reporting to their engineering manager, is in a different position from one who delivers three scoped pieces of work to three organisations in a year, whatever either contract says. If you intend to operate outside IR35, structure engagements so the day-to-day reality supports it.

The India mechanics: GST, TDS, moonlighting and cross-border clients

India has no IR35 equivalent, which Indian engineers sometimes read as meaning the independent route is simpler here. It is not; the complexity sits elsewhere — registration, indirect tax, withholding, and employment-contract issues with no UK counterpart.

The first decision is your vehicle. Most people start as a sole proprietor because it is the path of least resistance: a current account, a PAN, and you can invoice. As income grows the calculus changes. A private limited company or an LLP gives you limited liability, a clean separation between personal and business finances, better standing with overseas clients wary of paying individuals, and more structured ways to handle profit — against incorporation costs, annual filings, audit thresholds and a compliance calendar. Where that crossover sits depends on liability exposure as much as income; it is a conversation with a chartered accountant, not a rule of thumb.

The second is GST. Registration turns on turnover thresholds and the nature of your supply, and once registered you charge, collect, remit and file, with reconciliation. The structural point for AI engineers is that export of services to an overseas client is treated differently from a domestic invoice, with its own conditions and paperwork around the receipt of foreign currency. This is exactly where informal advice from other freelancers goes out of date, because the rules have been revised repeatedly. Check the official GST portal or ask an adviser.

The third is TDS. Indian clients typically deduct tax at source from professional fees and deposit it against your PAN, where it appears in your annual tax credit statement and is set against your final liability. Two consequences: your cash flow through the year is smaller than your invoices suggest, and you must reconcile what clients say they deducted against what actually appears. Discrepancies are common and far easier to fix within the same financial year. The income-tax portal is where you check.

Fourth, worth an explicit conversation with an adviser: India offers a presumptive taxation route that many independent professionals use, under which income is computed as a stated proportion of receipts rather than by tracking every expense, subject to eligibility and turnover limits. For an engineer whose costs are modest it can simplify life considerably. Whether you qualify, and whether it helps in your case, is what a professional's fee buys. This article quotes no rates or limits, because they move.

Then the two India-specific career issues. The first is moonlighting clauses. Indian employment contracts commonly contain exclusivity and non-compete language, and several large Indian IT services firms and global capability centres have treated undisclosed second jobs as a terminable offence rather than a grey area, including where the second engagement was unrelated. Plan around that: read the contract, and if you proceed, get written permission rather than assuming tolerance. An IP-assignment clause is a separate problem — permission to do the work is not ownership of what you produce.

The second is contracting for a US or UK client from India, now a substantial share of well-paid independent AI work. Three routes dominate. Direct invoicing is simplest and highest-margin: you invoice, they remit in foreign currency, your bank handles the inward remittance paperwork. A staffing intermediary takes a margin and handles contracting, payment risk and sometimes compliance. An employer-of-record has a third party employ you locally in India on the client's behalf — payroll, statutory benefits, PF — while the client gets a compliant engagement without an Indian entity. If you are also weighing physical relocation, our guide to visas and relocation between India and the UK covers that side.

Across all three routes, the thing Indian AI contractors most often get wrong is intellectual property. The deliverable is rarely just code: it is model weights produced by fine-tuning, the fine-tuning and data-preparation scripts, the evaluation datasets you constructed, the prompt assets, and the inference code. A contract assigning "all work product" without enumerating those artefacts leaves real ambiguity about who owns a fine-tuned checkpoint derived from client data using your pipeline. Name them, and make the boundary between client deliverables and your pre-existing tooling explicit too.

Contract clauses an AI engineer should read twice

The following applies in both markets, and to employment contracts as much as engagement letters — several of the worst clauses people sign are in permanent offers.

Two deserve more than a table row. Uncapped liability can end you financially, and it is increasingly common in AI contracts because clients are nervous about model behaviour. A cap at some multiple of fees paid is a normal ask, as is refusing to indemnify a client against the outputs of a model you did not train on data you did not choose.

And publicity rights are a career cost almost nobody prices. If you cannot say what you built, cannot describe it even anonymised, and cannot reference the outcome, a year of your working life is invisible to every future employer and client. A permanent employee survives that because tenure is legible on its own; for a contractor it is close to fatal, because your only asset is demonstrable work — the whole argument of our guide to building an AI engineering proof-of-work portfolio. If a client genuinely requires total silence that is their right, but it belongs in the rate, and you should say so during the commercial conversation.

What each model does to your career, two years out

Income is the axis everyone models. Career trajectory is the axis that compounds, and it is where the three differ most.

Permanent buys depth. Its unique property is that you are still there when your decisions produce consequences. You chose the vector store, and eighteen months later you are dealing with ten times the document count. You designed the evaluation harness, and you are there when it misses the regression that reached production. That loop — decide, live with it, learn — is what senior AI hiring rubrics probe for, and the substance behind our junior-to-staff AI career ladder. The failure mode is stagnation: excellent, narrow skills at an organisation whose problems you have stopped finding interesting. The signal to move is being able to predict the next twelve months accurately.

Contract buys breadth and pattern recognition. After six engagements you have watched six teams solve overlapping problems differently, and you develop a real instinct for which approaches survive production. You also learn commercial skills employees rarely acquire: scoping, pricing, saying no, reading an organisation quickly. The failure mode is a CV of shallow six-month engagements with no owned outcomes — arriving after the interesting decisions and leaving before the consequences. The signal to change is noticing you have not seen the second year of anything you built.

Fractional buys optionality, and is realistically viable only once you have a reputation someone will pay a premium for. Its distinctive fragility is concentration: a portfolio where one client is 70 per cent of your income is not a portfolio, it is a permanent job without the protections, and it can end at a single budget review. The other failure mode is dilution — three days across three clients often means being slightly present everywhere and consequential nowhere. The signal to consolidate is being unable to name a decision you owned last quarter.

Which leads to the point that justifies this article. At senior levels, demonstrated ownership of shipped outcomes matters more than the employment structure on your CV. Nobody interviewing for a staff role in Bengaluru or London awards points for having been permanent. They award points for "I built the evaluation infrastructure for a retrieval system serving four million queries a month, here is what it caught, here is what it missed, here is what I changed after the second incident." That story is equally tellable from all three models — and equally untellable from any of them if you never stayed for the consequences or never secured the right to describe them.

How to choose, this year

The framework is short, and deliberately so.

Early career and still building depth: prefer permanent. You need the consequence loop, senior people reviewing your work, and someone else absorbing the risk while you find out what you are good at. Choose the employer carefully — our checklist for vetting an AI start-up and our guide to equity, vesting and dilution are what to read before signing.

Deep specialism, a pipeline of inbound demand, and six-plus months of runway: contracting is a reasonable bet. All three, not two. Runway and utilisation are the preconditions; skill is assumed. If you have never had an enquiry you did not chase, you do not have a pipeline — build one from an employed position first, using our guide to landing your first AI consulting clients.

Established reputation and a preference for portfolio work: fractional is available to you. Set a concentration limit and hold to it even when the largest client offers more days. Keep a search discipline running whatever your model; our job-search operating system works as well for pipeline as for job hunting.

# ready.py — check the two preconditions instead of assuming them.

MONTHS_RUNWAY_MIN = 6
UTILISATION_MIN = 0.60           # billable days / available days
CONCENTRATION_MAX = 0.50         # share of income from one client


def months_of_runway(cash, monthly_costs):
    return cash / monthly_costs


def utilisation(billed_days, available_days):
    return billed_days / available_days


def concentration(revenue_by_client):
    total = sum(revenue_by_client.values())
    return max(revenue_by_client.values()) / total


def go_independent(cash, monthly_costs, billed, available, inbound_enquiries):
    checks = {
        "runway":      months_of_runway(cash, monthly_costs) >= MONTHS_RUNWAY_MIN,
        "utilisation": utilisation(billed, available) >= UTILISATION_MIN,
        "pipeline":    inbound_enquiries >= 3,
    }
    return all(checks.values()), checks


if __name__ == "__main__":
    print(go_independent(24000, 3000, 55, 90, 2))
    print("concentration:", concentration({"a": 60, "b": 25, "c": 15}))

Whichever you choose, the same thing decides whether it works: whether the people with budget can find you and see evidence of what you have done. Independents and fractionals live or die on inbound demand, and the difference between 140 billable days and 200 is rarely talent — it is visibility. A public profile showing verified work is how hiring managers and founders across India and the UK find specialists rather than generalists. Early profiles carry the Founding Builder badge, and those spots are limited.