By Moksh ParikhKey takeaways
- You cannot judge code, but you can judge evidence: live work you can open, a client you choose to call, and a test on your own data before the full build.
- A contract protects you only if it has a 'done when' line for each feature, a written change process, a signed IP assignment, your name on every account and a handover list.
- Fixed price and time-and-materials can both work; what makes either safe is short phases, a cap, written acceptance and the right to stop with everything you have paid for.
- For AI work, ask how it will be tested on your data, what it does when it does not know, which provider and paid plan it uses, and what it costs per use at your volume.
- Treat guaranteed accuracy, guaranteed returns or a promise to replace your staff as a reason to walk away or demand evidence.
Nearly every guide on how to choose an AI development company is written by an AI development company, and most cite nothing. You cannot read code, and you should not need to: what you can check is evidence, the contract, and a small paid test on your own data. Below are a 100-point vendor scorecard with the answers that mean walk away, 18 questions with a weak and a strong answer to each, and the clauses that keep you in control.
We are a vendor too, and we flag it where that matters.
Why projects go wrong, and what it means for your shortlist
The best evidence on overruns comes from big projects, so read it for its shape, not its numbers. McKinsey and the University of Oxford studied more than 5,400 IT projects with budgets above $15 million: on average they ran 45% over budget and delivered 56% less value than planned (McKinsey and Oxford, 2012). Flyvbjerg and Budzier found an average cost overrun of 27% across 1,471 IT projects, but one in six was a “black swan” averaging a 200% overrun (Harvard Business Review study).
A $40,000 build is not a $15 million programme, so do not borrow those percentages. Borrow the lesson: the danger sits in the tail, so cap your worst case by buying in short phases, seeing working software often and keeping the right to stop.
AI adds its own ways to fail. RAND interviewed 65 experienced data scientists and engineers and found five root causes: the problem is misunderstood or measured wrongly; the right data does not exist; the team chases new technology instead of the user’s problem; the systems to run the model are missing; or the problem is too hard for current AI (RAND’s root-cause study). RAND excluded projects that only used ready-made language models, as most small-business assistants do, so treat the list as a map of risks, not a failure rate.
Step 1: Shortlist with evidence, not badges
Review directories are useful if you know what the stars prove. Clutch says reviewers sign in with LinkedIn, Google or a company email and are checked for identity and work history; some reviews it cannot verify are published marked “Not Verified” (Clutch: how reviews are verified). But the vendor chooses which clients to invite (Clutch: asking clients for reviews), so reviews lean toward happy clients. Paying sponsors are listed above non-sponsors on directory pages, although Clutch says sponsorship does not affect its Leaders Matrix ranking (Clutch: sponsors and non-sponsors). That ranking gives reviews 20 of its 40 “Ability to Deliver” points (Clutch: ranking factors).
Fake reviews are banned under a 2024 FTC rule (FTC rule on fake reviews) and, in the UK, since 6 April 2025 (CMA fake reviews guidance). That removes fraud, not selection.
Use directories for a long list and for reviewers whose project size matches yours, and read for what went wrong. Then ask for the vendor’s last two clients, not its best two, and call one the vendor did not choose.
Step 2: Read the proposal like a contract
Most disputes start with a sentence nobody wrote down.
Scope, acceptance and change control
The Scrum Guide defines a “Definition of Done” as “a formal description of the state of the Increment when it meets the quality measures required for the product” (The Scrum Guide). The plain version: every feature in the proposal gets a “done when” line, such as “works on phone, tablet and desktop”. In 2019 Hertz sued Accenture for $32 million after a website and app rebuild missed its launch, alleging among other things that the tablet layout was left out and extra fees were demanded to add it; Accenture called the claims “without merit” (Consulting.us report). Unwritten requirements become change requests.
Add one change-control clause: every change is written down with its cost and time impact before work starts, and nothing is billed that you did not approve in writing.
Fixed price or time-and-materials
We usually quote a fixed price against a written scope, so weigh this section with that in mind. The strongest public argument points the other way. The US government’s 18F team advised agencies to buy custom software on time-and-materials “with a not-to-exceed ceiling”. Fixed price, it argued, assumes the requirements can be known up front, while time-and-materials “allows for much easier escape clauses”: if the work is poor, you stop assigning it. Its buyer questions mark “fixed price, because it’s the best way to control vendor costs” as the wrong answer (18F de-risking guides). That advice assumes a product owner who directs the work daily; most small firms lack one, and pure time-and-materials then leaves the hours to someone who may not be able to judge them.
| Model | Works best when | Main risk to you | Protection to ask for |
|---|---|---|---|
| Fixed price, whole project | Small, well-understood scope, measured in weeks | Price padded or corners cut; every change becomes a paid change request | Written acceptance criteria; a change-control clause; payments tied to accepted deliveries |
| Fixed price per phase, after paid discovery | Scope can be pinned down one phase at a time | Phases drift if discovery was rushed | Phases of 2–6 weeks; the right to stop after any phase; you own everything delivered so far |
| Time-and-materials with a cap | Scope is genuinely uncertain, such as a new AI use or messy data | Hours run on with nothing usable | A not-to-exceed ceiling; a working demo every week; timesheets; stop on notice |
| Dedicated developer or team, monthly | An ongoing roadmap you can direct | You manage the work, so quality depends on your oversight | Named people; a notice period; code in your repository every day |
The cap on time-and-materials follows the 18F guidance; the rest is our assessment. Checked 4 October 2026.
Safety does not come from the pricing model. It comes from short phases, written acceptance, a cap, and the right to walk away with what you have paid for. Ask which row each quote sits in; our software development page shows how we scope a build.
Who owns the code, the accounts and the data
Paying for code does not make it yours. In the US, contractor work is “made for hire” only if it fits one of nine listed categories under a signed agreement, and software is not one of them (17 U.S.C. §101). A transfer of ownership “is not valid unless” it is written and signed (17 U.S.C. §204), so you need a written assignment, not just the words “work for hire”. In the UK, the creator owns commissioned work “unless you otherwise agree it in writing”, and without a clause you may get only a limited licence to use it (UK Intellectual Property Office). We have not covered Indian law here. This is general information, not legal advice; have a lawyer read any contract that matters.
Then make ownership real: the code repository in your own account from day one, with the vendor as a user (the 18F guides ask for work to be committed at least daily); cloud, domain, app-store, payment and AI-provider accounts in your business’s name; and a list of any pre-existing parts the vendor keeps, with your licence to use them.
Data protection and security you can test
If the vendor handles personal data, some terms are legal requirements. Under UK GDPR Article 28, the contract must cover your documented instructions, staff confidentiality, security, no sub-processors without your written permission, help with rights requests and breaches, deleting or returning the data at the end, and audits (ICO: what must be in the contract). California requires similar written terms with service providers (11 CCR §7051) for businesses above its thresholds, such as revenue over $26,625,000 (California Privacy Protection Agency). India’s data protection rules, notified on 14 November 2025, allow penalties of up to ₹250 crore for failing to keep reasonable security safeguards (PIB backgrounder). On an AI project, the sub-processor list should name the model provider.
“We follow best practice” cannot be tested. OWASP calls its Top 10 “a standard awareness document” (OWASP Top 10:2025), while its ASVS 5.0.0 “provides a basis for testing web application technical security controls” (OWASP ASVS). Ask for “built and tested to ASVS Level 1”, or Level 2 for sensitive or payment data. In the UK, Cyber Essentials is the government-backed baseline, from £320 plus VAT, with certificates you can look up (NCSC on Cyber Essentials). Few buyers check: only 15% of UK businesses had reviewed the cyber risks from their immediate suppliers (DSIT breaches survey).
Exit and handover
Plan the ending first. The UK government’s AI Playbook tells buyers to write “strategies to avoid vendor lock-in” into their specifications (AI Playbook for the UK Government), and 18F’s right answer for how long it should take to replace a failing vendor was about six weeks, not “months or years”. Ask for a written exit list: source code, setup notes, credentials, data exports, AI prompts and test sets, plus a day rate for handover help.
Step 3: The AI checks most buyers skip
A demo on the vendor’s own examples proves little. Five checks prove more.
Test on your data first. Anthropic says tests should “mirror your real-world task distribution”, edge cases included (Anthropic: define success criteria); OpenAI recommends mixing real data with expert-built test cases and re-testing on every change (OpenAI evaluation best practices). The FTC’s DoNotPay order, with a $193,000 payment, turned on an “AI lawyer” never tested against a human lawyer’s work (FTC order against DoNotPay). Agree a test set from your real questions, a pass mark and who grades it. Our suggestion, not a standard: 50 to 100 real customer questions, with the answers written by your staff.
Ask what it does when it does not know. Anthropic recommends letting the model say “I don’t know”, requiring citations and limiting answers to the documents supplied, and warns that these steps reduce made-up answers but “don’t eliminate” them (Anthropic: reduce hallucinations). A Canadian tribunal held Air Canada liable for wrong refund advice from its website chatbot, finding the company “is responsible for all the information on its website” (McCarthy Tétrault on Moffatt v. Air Canada). Test that behaviour before launch.
Ask which provider and which plan. OpenAI says API data has not been used for training since 1 March 2023 unless you opt in (OpenAI: your data), and Anthropic’s commercial terms say it “may not train models on Customer Content” (Anthropic commercial terms). Google’s terms say content on the unpaid Gemini API tier is used to improve its products and may be read by human reviewers; paid-tier content is not (Gemini API terms). Get the provider, plan and data region in writing.
Ask the cost per use. Models charge per token, a small chunk of text, at list prices like these.
| Model | Input, per million tokens | Output, per million tokens |
|---|---|---|
| Anthropic Haiku 4.5 | $1 | $5 |
| Anthropic Sonnet 5.5 | $2 | $10 |
| Anthropic Opus 5.5 | $4 | $20 |
| OpenAI gpt-6-luna | $0.10 | $0.50 |
| OpenAI gpt-6.1-sol | $2 | $10 |
Sources: Claude pricing and OpenAI API pricing, checked 4 October 2026.
Worked example. A document assistant. Assumptions (ours): each answer sends 4,000 input tokens and returns 500 output tokens, 3,000 times a month. On Sonnet 5.5: 4,000 × $2 ÷ 1,000,000 = $0.008, plus 500 × $10 ÷ 1,000,000 = $0.005, so $0.013 per answer, and $0.013 × 3,000 = $39 a month. On Haiku 4.5 it is $0.0065 per answer, or $19.50 a month; on Opus 5.5, $0.026, or $78. Model fees only, before hosting, monitoring and support.
A vendor should give you this formula at your volume, say what makes it rise, and set caps and alerts; OWASP lists “unbounded consumption” among the top risks for AI applications (OWASP Top 10 for LLM Applications).
Ask how quality is watched, and distrust guarantees. NIST’s AI Risk Management Framework puts ongoing monitoring under its “Manage” function (NIST AI RMF): ask what is logged, who reviews it, and how a drop is caught when the provider updates its model. The FTC said Workado advertised 98% accuracy when testing showed 53% on general content (FTC order against Workado), and the FTC has sued Air AI, alleging it sold AI to small businesses as able to replace staff, with some losing up to $250,000 (FTC complaint against Air AI). Treat any guaranteed accuracy or return as a cue to ask for evidence on your data.
This is the order we follow for AI automation: a test on your data first, then the build.
Step 4: Run a small paid pilot before you commit
UK government teams start with a discovery phase, where “around 4 to 8 weeks is typical” and “it’s not a failure to stop at the end” (GOV.UK: how discovery works). An alpha of six to eight weeks then tests the riskiest ideas with throwaway prototypes (GOV.UK: how alpha works). A small business can shrink that into a two-to-six-week paid pilot:
- Pick one workflow with a pain you can measure.
- Write success criteria first: correct, sourced answers to at least X of 100 real questions; “I don’t know” for anything outside your documents; under $Z per 1,000 answers. You set X and Z; no outside benchmark exists.
- Use real data, on a paid AI plan, under an NDA and data-processing terms.
- Fix the fee. You keep the code, prompts, test set and results whatever you decide.
- Test the riskiest part first: data quality or the key integration, not the nicest screen.
- Decide at a gate: stop, repeat, or move to a fixed-price or capped build. Write the stop condition down in advance.
- Judge the vendor, not the demo. Did they flag problems early and tell you what would not work?
We found no independent market price for a pilot; the figures online are vendors quoting themselves.
India or onshore: cost, working hours and risk
Disclosure: our team is based in Ahmedabad, India, so we have a stake in this answer.
| Source | Who | Rate |
|---|---|---|
| Clutch pricing guide | Agencies in India | $25–$49 an hour, the most common band |
| Clutch pricing guide | Agencies in the United States | $50–$99 an hour, the most common band |
| Clutch pricing guide | Agencies in the United Kingdom | listed as “Unknown” |
| Accelerance survey of 60 partners | South Asia | $24–$31 an hour junior; $31–$41 senior |
| Accelerance survey of 60 partners | Central and Eastern Europe | $64–$76 an hour senior |
| Accelerance survey of 60 partners | Latin America | $60–$75 an hour senior |
| US Bureau of Labor Statistics | US software developers, employed | median $135,980 a year |
| IT Jobs Watch | UK freelance software developers | median £525 a day |
Sources: Clutch pricing guide, updated 21 Sep 2026; Accelerance 2026 outsourcing rates, 24 Nov 2025; BLS, May 2025 data; IT Jobs Watch, six months to 4 Oct 2026. Checked 4 October 2026.
Accelerance itself warns that hourly rates are a poor measure of true cost.
Worked example. 400 hours of build work, about ten weeks for one developer. Indian agency: 400 × $25 = $10,000, up to 400 × $49 = $19,600. US agency: 400 × $50 = $20,000, up to 400 × $99 = $39,600. UK freelancer at the median: 400 ÷ 8 = 50 days, and 50 × £525 = £26,250, before the project management an agency would add. Assumption: equal hours mean equal output, which is not guaranteed.
India Standard Time is UTC+5:30 with no daylight saving (Britannica), and UK clocks go back on 25 October 2026 (GOV.UK clock changes). Our overlap calculation, for offices open 9:00 to 17:30:
| Office | Indian team’s hours | Overlap in summer | Overlap in winter |
|---|---|---|---|
| London | 9:30–18:30 IST | about 5 hours | about 4 hours |
| New York | 9:30–18:30 IST | none | none |
| New York | 13:00–22:00 IST | 3.5 hours | 2.5 hours |
US West Coast overlap is close to nil without overnight shifts, so ask how a vendor staffs your hours.
India is not on the UK’s adequacy list, so a UK firm sending personal data there generally needs the ICO’s International Data Transfer Agreement or the UK Addendum, plus a transfer risk assessment (DLA Piper on UK data transfers). Keep payments staged and ownership in your own accounts, and you will rarely need to enforce a contract across borders. With any small vendor, local or offshore, ask who picks up the work if one key person leaves. Offshore suits you less if the work needs daily in-person contact, the data cannot leave the country, or nobody can review progress weekly. If you want developers on your hours, under your direction, see how our dedicated developers work.
The vendor scorecard
Score each vendor from 0 to 5 on each row, multiply by the row’s weight and divide by 5. The weights are our judgement, drawn from the risks above.
Walk away, whatever the score, if the vendor:
- won’t assign the IP to you in writing;
- won’t put the scope and acceptance criteria in writing;
- insists on holding your code, cloud or domain accounts;
- guarantees accuracy or a return on investment;
- won’t sign data-processing terms;
- can’t name who will do the work;
- can’t give you one reference you can call.
| # | Criterion | Weight | 0–1 looks like | 4–5 looks like | Evidence to ask for |
|---|---|---|---|---|---|
| 1 | Proof of similar, live work | 20 | Mock-ups only; “it’s all under NDA” | A live product of similar size; two recent clients you can call, one of your choosing | Links, reference calls, reviews from projects of a similar budget |
| 2 | Scope, price and change control | 15 | A one-line quote; “we’re agile, we’ll see” | Written scope with “done when” lines; phase prices or capped time-and-materials; a written change process | Draft statement of work; a sample change request |
| 3 | Ownership and exit | 15 | Vendor-held accounts; a vague IP clause | Your repository and accounts from day one; an assignment clause; a handover list and rate | The contract clauses; an invitation to your repository |
| 4 | Security and data protection | 15 | “We’re secure” | A named standard (OWASP ASVS Level 1 or 2); Cyber Essentials or SOC 2 where relevant; Article 28 or California service-provider terms; a sub-processor list | Certificates, data-processing agreement, sub-processor list |
| 5 | AI competence and honesty | 15 | Demo only; “no hallucinations” | A test on your data with an agreed pass mark; “I don’t know” handling; named provider and paid plan; cost-per-use formula; monitoring plan | Pilot proposal, test plan, cost sheet |
| 6 | Team and communication | 10 | Only the salesperson; no plan for shared hours | Named team met before signing; weekly demo; fixed overlap hours | Team list; a sample weekly report |
| 7 | Total cost, build plus running | 10 | Low rate, unclear extras | Full 12-month cost, including hosting, AI usage and support | A 12-month cost sheet |
| Total | 100 |
No AI in the project? Drop row 5 and spread its 15 points across rows 1 to 4.
Worked example. A vendor scores 4, 3, 5, 4, 3, 4 and 3 on rows 1 to 7. Weighted: 20 × 4 ÷ 5 = 16; 15 × 3 ÷ 5 = 9; 15 × 5 ÷ 5 = 15; 15 × 4 ÷ 5 = 12; 15 × 3 ÷ 5 = 9; 10 × 4 ÷ 5 = 8; 10 × 3 ÷ 5 = 6. Total: 75. The weak rows, scope and AI, can both be fixed in the contract before you sign.
Reading the total: 80 or more, go to a paid pilot. From 65 to 79, go ahead only if the gaps can be fixed in the contract. Under 65, keep looking. The same rules apply to us.
18 questions to ask a software development company before you sign
The format borrows from 18F’s “wrong answer, right answer” buyer questions; the questions are ours. They suit any software build, and questions 12 to 16 are specific to AI.
| # | Question | Weak answer | Strong answer |
|---|---|---|---|
| 1 | Show me something you built that is live today. Can I call that client? | Screenshots, and every client is “under NDA” | A live link and a recent client you can phone |
| 2 | Who exactly will do the work, and can I meet them before signing? | “Our team” | Names, and a call with them this week |
| 3 | What will I see working in the first two weeks? | “Nothing until the design is signed off” | A small, specific piece you can click |
| 4 | For each feature, how will we both know it is done? | “We’ll know it when we see it” | A “done when” line for every feature |
| 5 | Fixed price or time-and-materials, and why for this project? | The same model for every client | The trade-off explained; phase prices or a cap |
| 6 | What happens if I change my mind halfway through? | “No problem” (then an invoice) | A written change request with cost and time, approved before work starts |
| 7 | Whose name is on the code repository, cloud, domain and AI accounts? | “Ours, it’s easier” | “Yours from day one; we are users” |
| 8 | Where in the contract do you assign the IP to us, and when? | “Don’t worry, you’ll own it” | Points to a signed assignment clause |
| 9 | If we part ways, what do we get, and how long would a new team need? | “You’d have to start again” | A handover list; weeks, not months |
| 10 | Which security standard will you build and test against? | “Industry best practice” | “OWASP ASVS Level 1 (or 2), tested before launch” |
| 11 | Will you sign a data-processing agreement, and who are your sub-processors? | “We don’t need one” | Yes, with a list that includes the AI provider |
| 12 | Which AI provider and plan will you use, and is our data used for training? | “Don’t worry, it’s private” | Named provider, paid plan, data region and training terms |
| 13 | How will you prove it works on our data before the full build? | A polished demo on their own examples | A test set from your data, an agreed pass mark, a named grader |
| 14 | What does it do when it doesn’t know the answer? | “Our model doesn’t hallucinate” | Says so, cites its sources, hands off to a person |
| 15 | What will it cost each month at our volume, and what stops a runaway bill? | “Very little” | A cost-per-use formula, caps and alerts |
| 16 | How will we know if quality drops after launch? | “Call us if something looks wrong” | Logging, regular test runs and a named owner |
| 17 | Tell me about a project that went wrong and what you changed. | “That’s never happened to us” | A specific story and the fix |
| 18 | What would make you tell us not to build this? | “We can build anything” | Clear conditions: no usable data, an unclear problem, or a task too hard for AI |
Red flags and green flags
| Red flag | Green flag |
|---|---|
| Quotes a price within a day without asking about your process or data | Suggests a short paid discovery or pilot first |
| Promises guaranteed accuracy, a guaranteed return or “replaces your staff” | States the limits, and what the AI should not do |
| A feature list with no acceptance criteria | A “done when” line for each feature |
| Code, cloud or domain in the vendor’s name; a vague IP clause | Your accounts from day one; a written assignment |
| Can’t say who will do the work | A named team you meet before signing |
| “We’re secure”, with no standard named | An OWASP ASVS level named; Cyber Essentials in the UK |
| A prototype on a free AI tier using real customer data | Named provider, paid plan and data region, on the sub-processor list |
| Exit means “start again” | A written handover list and day rate |
Sometimes the right answer is not to hire anyone yet. If you cannot describe the problem in two sentences, the data does not exist, or nobody can spare an hour a week to review progress, any vendor will spend your money finding that out. Fix those first: write the problem down, collect 50 real examples, and name the person who will own the result. Then use the scorecard, ask the 18 questions, and buy the first phase small.
Questions people ask
What is the single best question to ask an AI development company?
Ask them to show it working on your own data, against success criteria you both agree in writing, before you pay for the full build. Anthropic and OpenAI both tell developers to test on examples that mirror real use, including awkward edge cases. A vendor who will only demo on their own examples has not yet shown you anything.
Should I choose fixed price or time-and-materials?
It depends on how well the scope is known, and both can work. The US government's 18F team recommended time-and-materials with a not-to-exceed cap for custom software, because requirements change and it is easy to stop; fixed price suits small, well-defined phases, especially when nobody on your side can manage the hours day to day. Whichever you choose, insist on short phases, written acceptance criteria, a spending limit and the right to stop.
Do I automatically own the code I pay a developer to write?
No. In the US, software from an outside contractor does not fall into the categories that can be made a work for hire by contract, so ownership passes only through a signed written assignment. In the UK, the creator owns commissioned work unless you agree otherwise in writing. Put an assignment clause in the contract, keep the code in your own account, and have a lawyer check any contract that matters.
Are Clutch reviews real?
Clutch checks reviewers' identity and work history, and labels reviews it cannot verify as Not Verified. But vendors choose which clients to invite, and paid sponsors appear above non-sponsors on directory pages, although Clutch says sponsorship does not affect its Leaders Matrix ranking. Use reviews to build a shortlist, then call a recent client the vendor did not pick.
How do I stop an AI assistant from making things up?
You cannot stop it completely, but you can reduce and contain it. Limit answers to your own documents, make it show its sources, let it say it does not know, and hand hard questions to a person. Test this before launch, because a Canadian tribunal held Air Canada responsible for wrong refund advice its website chatbot gave a customer.
Will my business data be used to train the AI model?
On the main paid API plans, generally not: OpenAI says API data is not used for training unless you opt in, and Anthropic's commercial terms bar it from training on customer content. Free tiers can differ; Google says content on the unpaid Gemini API tier is used to improve its products and may be read by human reviewers. Ask your vendor which provider and plan they will use, and get it in writing.
Is it cheaper to hire an Indian development company?
Per hour, usually: Clutch's pricing guide shows Indian agencies most often charging $25 to $49 an hour, against $50 to $99 for US agencies. The saving shrinks if the scope is vague or nobody reviews the work each week. US clients also get little overlap in working hours unless the Indian team shifts its day.
What does it cost to run an AI feature each month?
It depends on volume and the model. On our assumptions, a document assistant answering 3,000 questions a month costs roughly $20 to $80 in model fees at Anthropic's prices checked in October 2026, before hosting and support. Ask your vendor for the cost-per-use formula at your volume, and for spending caps and alerts.
Sources
- 18F — De-risking Government Technology guides (archive) · last updated Apr 2024
- RAND — The Root Causes of Failure for Artificial Intelligence Projects · 13 Aug 2024
- Harvard Business Review (arXiv abstract) — Why Your IT Project May Be Riskier Than You Think · Sep 2011
- Clutch Help Center — How Clutch verifies reviews · accessed 4 Oct 2026
- Cornell Legal Information Institute — 17 U.S.C. §204 (transfers of copyright must be in writing) · current statute
- GOV.UK (Intellectual Property Office) — Ownership of copyright works · 19 Aug 2014
- Information Commissioner's Office — What needs to be included in the contract? · accessed 4 Oct 2026
- OWASP — Application Security Verification Standard · v5.0.0, May 2025
- Department for Science, Innovation and Technology — Cyber security breaches survey 2025/2026 · 30 Apr 2026
- Anthropic docs — Reduce hallucinations · accessed 4 Oct 2026
- OpenAI docs — Evaluation best practices · accessed 4 Oct 2026
- US Federal Trade Commission — FTC finalizes order with DoNotPay that prohibits deceptive AI lawyer claims · Feb 2025
- McCarthy Tétrault — Moffatt v. Air Canada: A Misrepresentation by an AI Chatbot · 19 Feb 2024
- Anthropic — Claude pricing · checked 4 Oct 2026
- Clutch — Software Development Company Pricing Guide · updated 21 Sep 2026
- GOV.UK Service Manual — How the discovery phase works · 21 Jun 2021
How we research: every price in this article was checked on the vendor's own page, and every claim links to where it came from, as of 4 October 2026. Prices change — confirm them before you buy.