Gartner placed agentic AI at the Peak of Inflated Expectations in its first standalone hype cycle for the category in April 2026, with only 17% of organizations having deployed an agent and more than 40% of agentic projects expected to be canceled by the end of 2027 (Gartner). Every one of those canceled projects had a vendor. Most of those vendors had a polished demo, a confident deck, and no answer to the questions in this post.
We are an AI agent development company, so read this with that in mind. It is also why we can tell you exactly which questions make weak vendors uncomfortable: we get asked them, and we have watched competitors dodge them. This post gives you the 12 questions, the answers you should hear, the pricing shapes that are fair, and the red flags that predict a rollback. It does not rank vendors; the third-party directories do that, and we would rather you judge the work than the listing.
What does an AI agent development company actually do?
An AI agent development company designs, builds, integrates, evaluates, and operates AI systems that take actions inside your business, such as resolving support tickets, processing documents, or qualifying leads, using a language model connected to your tools and data. The work is mostly integration, evaluation, and security engineering; the model is a component, not the product. That distinction is the first thing to test a vendor on, because companies that sell "the AI" rather than "the system" are the ones whose agents get rolled back.
Why vendor selection matters more for agents than for ordinary software
Three 2026 data points explain why the vendor decision carries more risk here than in a normal build. Sinch found 74% of enterprises had rolled back a live AI communications agent, and the rate was higher, 81%, among companies with mature governance, because their monitoring caught failures others missed (Sinch). Menlo Ventures reported that 76% of enterprise AI use cases were bought rather than built in 2025, up from 53% (Menlo Ventures), which means most buyers are choosing partners, not writing code. And in May 2026 Anthropic, Blackstone, Hellman & Friedman and Goldman Sachs launched a $1.5 billion services company aimed squarely at mid-sized firms (CNBC). When the model makers start selling implementation, the implementation is where the value and the risk live.
The market has also filled with newcomers. "AI automation agency" is a 2,900-search-a-month term in the US and many of those agencies were founded in the last 18 months by people who have configured a no-code tool and never carried a production system through its first outage. That is not disqualifying on its own. It is disqualifying when combined with the answers below.
The 12 questions, and the answers that separate builders from demo shops
| # | Question | Good answer | Red flag |
|---|---|---|---|
| 1 | Show me an agent you built that has been in production for six months. What broke? | Named client or anonymized specifics, a real incident, and what changed after it | Only demos, or "nothing broke" |
| 2 | How will we measure whether this worked? | A baseline metric, a holdout, and a payback calculation before the SOW is signed | "Time saved" with no unit cost |
| 3 | What will this cost to run per month at our volume? | A per-task token estimate from a pilot, plus fixed hosting and oversight | Build price only; run cost "depends" |
| 4 | What can the agent do without a human approving it? | A written capability list and a permission model, with irreversible actions gated | "It can do whatever you need" |
| 5 | How do you test it before a release? | An evaluation suite with a golden dataset, injection tests, and pass thresholds | "We test it manually" |
| 6 | What happens when the model provider changes prices, deprecates a model, or goes down? | Model abstraction layer, tested fallback, documented switch cost | Hard-wired to one vendor's proprietary builder |
| 7 | Who owns the code, prompts, evals, and data? | You, on payment, in writing, including the evaluation dataset | Prompts or "platform" retained by the vendor |
| 8 | How do you handle our data, and where does it go? | Named providers, regions, retention, a DPA, and zero-training clauses | Vague "enterprise-grade security" |
| 9 | Who is on the team, and who did the last three builds? | Named engineers with agent work in the last year, on this project | Senior faces in the pitch, unnamed juniors on delivery |
| 10 | What would make you tell us not to build an agent? | A clear answer: fixed-order workflows, low volume, no unit cost to remove | Nothing; every problem is an agent |
| 11 | What does support look like after launch? | Retainer with response times, monitoring, monthly eval review, prompt and tool updates | "Warranty period" then hourly |
| 12 | Can we talk to a client whose project you rolled back or scaled down? | Yes, and here is what we learned | No such client exists |
Question 6 became urgent this year. OpenAI announced it will sunset its no-code Agent Builder and hosted Evals on 30 November 2026, pointing customers to its Agents SDK instead (MCP Directory). Any vendor who built your agent on a proprietary builder just handed you a migration project. Ask where the logic lives, and insist that it lives in code you own.
What should an AI agent proposal cost?
Published 2026 ranges put a prototype at $10k to $30k, an MVP agent at $20k to $60k, and complex multi-system agents at $100k to $500k or more, with ongoing operations at 15% to 25% of build cost per year (Neontri). Our own tiers sit inside those bands and are itemized in our AI agent development cost guide. The number matters less than the shape. A fair proposal separates discovery, build, evaluation, and operation, and prices each.
| Proposal component | Fair structure | Warning sign |
|---|---|---|
| Discovery and scoping | Fixed fee, 1 to 3 weeks, ends with a written spec, a baseline metric, and a go or no-go | Skipped, or folded into a large fixed price |
| Build | Fixed price against the spec, or capped time and materials with milestones | Open-ended hourly with no cap |
| Evaluation and hardening | Explicit line item: golden dataset, eval suite, injection tests, security review | Absent, which means it will not happen |
| Run cost estimate | Tokens, hosting, monitoring, oversight, from pilot data | Missing or "pass-through" |
| Operation and support | Monthly retainer with named response times and a monthly eval review | Support only on request, at hourly rates |
| IP and data terms | Full assignment on payment, DPA, no training on your data | Vendor retains "framework" or prompts |
If a proposal is dramatically cheaper than the others, find the missing row. It is usually evaluation and hardening, and it is the row that determines whether you join the 74% who roll back. Our software development contract checklist covers the clauses that protect you once you have picked, and our RFP template makes the twelve questions above part of the formal process.
How to run the evaluation in three weeks
- Week 1: shortlist on evidence, not listings. Ask three to five vendors for one production case study each, with a number attached, and the answers to questions 1, 2, and 10 in writing. Drop anyone who cannot answer 10.
- Week 2: paid discovery with two vendors. A one-week fixed-fee scoping engagement with each, ending in a spec, a baseline metric, and a run-cost estimate. You are buying two independent opinions on your own problem for a few thousand dollars, and you keep both documents.
- Week 3: compare the specs, not the pitches. The vendor whose spec has the narrower scope, the clearer permission model, and the more conservative run-cost estimate is usually the one who has done this before. Check references, including question 12, then sign.
Red flags that predict a rollback
- The demo is the deliverable. A live demo on your data takes an afternoon. A production system takes an evaluation harness, permission design, and monitoring. Vendors who lead with the demo and cannot describe the rest are selling the afternoon.
- Multi-agent by default. Five named agents in an org chart is a conference architecture. Ask why one well-tooled agent would not do; if the vendor cannot answer, they have not read the cost data. Our analysis of when swarms earn their cost is in multi-agent systems for business.
- No security conversation until you raise it. A vendor who reaches the SOW without discussing least-privilege tools and approval gates has never been through an incident. The controls are in our AI agent security checklist; ask which ones the proposal includes.
- Everything is an agent. Fixed-order pipelines are workflows. A vendor who cannot recommend plain automation when it fits is optimizing for the invoice. The decision rules are in chatbot vs agent vs automation.
- ROI stated in hours saved. Hours are not money. Insist on a unit-cost model; the method is in how to measure AI ROI.
What experience looks like on our side of the table
We answer our own question 1 with TracefyHR, an HR SaaS platform we built and operate that serves 50+ companies and 2,000+ employees, including its Forge AI feature that generates working functionality from natural-language requests. The relevant part for a buyer is not the feature. It is that the system has been in production long enough to have a rollback history, an evaluation suite that grew out of real failures, and run-cost data we can show you. That is the bar to hold every vendor to, including us. The broader landscape of what AI development services include, and what each type costs, is in our AI development services buyer's guide.
Frequently asked questions
How much does it cost to hire an AI agent development company?
Published 2026 ranges run from $10k to $30k for a prototype, $20k to $60k for a production MVP agent, and $100k to $500k or more for complex multi-system agents, plus 15% to 25% of build cost per year to operate. Integration with your existing systems, not the model, is the largest cost driver.
Should we buy an agent platform instead of hiring a developer?
Buy when a packaged product already fits your workflow and your volume is modest; 76% of enterprise AI use cases are bought. Build with a partner when the agent must act inside your own systems, when data handling is sensitive, or when platform lock-in is a risk, as the November 2026 sunset of OpenAI's Agent Builder demonstrated. Many good outcomes are hybrids: a bought model, custom integration and evaluation.
What is the difference between an AI automation agency and an AI development company?
Automation agencies typically configure no-code and low-code tools to connect existing SaaS products, which suits simple, fixed-order workflows. AI development companies write and own code, build evaluation and permission systems, and can operate agents that take actions inside custom or regulated systems. Ask which one your problem needs before comparing prices.
How long does it take to build an AI agent?
A scoped single agent with two or three integrations typically takes 6 to 12 weeks including evaluation and a supervised pilot. Timelines under four weeks usually mean evaluation and hardening were skipped, which is the leading cause of post-launch rollbacks.
What should we own at the end of the project?
The source code, prompts, tool definitions, evaluation dataset and test suite, infrastructure configuration, and documentation, assigned to you on payment. If the vendor retains any of these, you do not own the agent; you rent it.
Key takeaways
- Agents are at the peak of the hype cycle. Only 17% of organizations have deployed one, and over 40% of projects are expected to be canceled. The vendor decides which group you join.
- Test for production experience, a measurement plan, run-cost estimates, and a permission model. Demo shops fail all four.
- A fair proposal separates discovery, build, evaluation, and operation. The cheapest quote is usually missing evaluation.
- Insist on owning code, prompts, and eval data, with a model abstraction layer. Proprietary builders can be sunset with months of notice.
- Run a three-week process: evidence-based shortlist, paid discovery with two vendors, compare specs rather than pitches.
- Any vendor who cannot name a problem that should not be an agent is optimizing for the invoice.
Put us through the twelve questions
If you are shortlisting AI agent development companies, send us your brief and we will answer all twelve questions above in writing within 48 hours, including run-cost estimates and a permission model for your specific workflow. If the honest answer to question 10 is that your problem is a workflow rather than an agent, we will tell you and quote the smaller build. That is the standard our AI development services are measured against.
Sources
- Gartner: Hype Cycle for Agentic AI (April 2026)
- Gartner: Over 40% of agentic AI projects will be canceled by end of 2027
- Sinch: The AI Production Paradox
- Menlo Ventures: The State of Generative AI in the Enterprise
- CNBC: Anthropic, Blackstone and Goldman Sachs AI services venture
- MCP Directory: OpenAI Agent Builder deprecation timeline
- Neontri: AI agent development cost in 2026