How long does it take to build an AI SDR in-house?

A few weeks, if you build on top of something that already exists. From a blank page, plan in months. Mine took 8 months of 2026, and I had 3 years of building agents behind me when I started. The market answer is "a weekend with a workflow tool", and that answer is measuring a different starting point without saying so.

I am Eduard Klein. I advise CEOs, founders and boards on AI Business Strategy, and the 8 months is my own number for the system in the full AI SDR blueprint, published 19 August 2026.

The two numbers, and why they are both true

Against an existing base I stand up a new agent with new capabilities in about 1 hour. I have done it repeatedly.

That is not a boast, it is an artefact of architecture. There is a service layer underneath that all my agents share. A research agent already hardened against prompt injection. An outreach agent that already handles sending, warmup control, inbox management. Nobody has to reinvent those, and Component 5 and Component 4 were the two that took longest to get right. Measured against that base, a new SDR is a matter of days.

From nothing, the honest figure is 6 to 8 months. That is what I paid in 2026 to build the base itself. I am not a beginner.

Both numbers describe the same system, the one Google's AI Overview and every vendor page are quoting at you from different ends. That is why the published ranges look absurd next to each other. "A weekend with a workflow tool" and "two quarters with a team" are both accurate, and neither one tells you which starting point it assumed. The range is not a disagreement about difficulty. It is a missing variable.

Why 8 months and not 8 weeks

The time did not go into writing code. I built, then I watched what I had built fail.

I built a version, I ran it, and I watched it break in ways I had not predicted. Then I rebuilt. Agent systems need longer test cycles than conventional software and constant correction, because the failure modes do not announce themselves in a unit test the way a Python exception does. In August 2026 I could finally describe the architecture with confidence, and the confidence came from the failures rather than the design.

The expensive line item in an AI SDR build is testing and adaptation. Not development. Every plan I have seen from Clay-plus-a-sequencer arithmetic onward has priced the writing and omitted the watching.

The variable is you, not the tooling

Build time is almost entirely a function of the builder's development experience. Not of the tools.

The tools in September 2026 are good. Claude Code and Codex will write most of the actual code. What they will not do is notice that a component has more access than its job requires, or that an agent keeping its own state has just made the pipeline unauditable. I made both of those mistakes before I wrote the 3 rules that now sit above every component, and finding them cost weeks each time.

A non-technical builder does not get a slower version of this timeline. They get a different one. I have watched it repeatedly since 2023: 3 months of something impressive running, then the agent restructures itself and breaks itself, and there is nobody in the room who can tell whether the fix is 2 days or a rewrite.

What you can compress and what you cannot

You can compress the code. You cannot compress the correction. I measured that across 3 years: Claude Code shortened the first half of that sentence and did nothing at all to the second.

The part that shrinks with better tooling is the writing, and between 2023 and 2026 it shrank by more than I expected, first with GPT and then with Claude. The part that does not shrink is the loop where you run the thing against reality, find out what it gets wrong, and feed the correction back in. In my system every correction I make is stored and reused as an example, which means the loop pays compounding interest. It still has to run for months before the output sits where I want it.

That is the honest answer to a founder asking for a date. If somebody promises you a working AI SDR in 3 weeks from a blank page, ask them what they are counting, and ask them how many months of correction they have run.

If you would rather not spend the 8 months, the alternative is to start from a framework that already exists. That is the build-or-buy decision in a different shape, and it is the work I do in AI coaching. The wider question of whether Google and ChatGPT can see your company at all is a separate piece: AI strategic visibility.

The number nobody asks for: when it is finished

It is not finished. That is the point rather than the problem.

My system has been permanently under development since 2023. You want it to fit your needs exactly, and your needs move, because markets move and products change. Previously none of that lived inside the software. You worked on the software from outside, and there was a release date.

For an AI SDR to make sense, all of that has to go into the agent system. That is the work, and it was always the work of an SDR. The difference in 2026 is that you do it with the agent rather than to a piece of software.

So the honest project plan has 8 months to a system you trust, and then no end date. If your board needs a completion milestone, that is a conversation to have before the build starts rather than in month 9.

What I still do not know

I do not know whether the 8 months holds for the next person who starts now.

The tools improved measurably in each of the 3 years I have been building on them. If Claude Code and Codex get better at structural judgement, the number falls, and it falls a lot. If they do not, the number is stable and the constraint stays human. I have watched for 3 years and I have not seen the structural judgement arrive yet. I will put the new figure on this page when I have measured it rather than guessed it.

What the timeline depends on: whether to build your own AI SDR or buy one, what an AI SDR actually costs and the tech stack I actually run.

About the author

Eduard Klein is an AI Business Strategist for CEOs, founders & boards, and he has watched software timelines from inside both the build and the sale: 30 years building software, plus software sales for SAP, for Microsoft, for Oracle and for the German Stock Exchange, where the gap between a promised date and a shipped one was somebody else's problem to explain. Years as an SDR up to 3.5 million in revenue, in key account management and as a sales manager in the software industry sit behind the estimates here, and he has published 170+ articles across 100+ publications including Cloudways and Userpilot (the full list).