AI leadership and delivery
For software companies and agencies with 15 to 200 people
Your engineers use AI. Make your delivery run on it.
Cone Red leads and builds the AI workflows behind your delivery, such as code review checks, tests, estimates and status reports, inside your systems. Dima Levin leads, our engineers build, and your people learn to run what we ship.
Send one workflow and three short answers. You'll get one page back within three business days. No call, no access to your systems, yours to keep.
The offer, start to finish
Our agents wrote and animated this film. The voice is Dima's, generated by AI from his recordings. Music: Dima's own track.
Where AI stalls
Your engineers write code faster. The slow part moved to review, testing and the client.
You pay for that wait out of your margin, or your roadmap slips. Six places it shows up, each with what we'd build and the effect we'd watch.
Writing and testing code accounts for about 25% to 35% of the time from initial idea to product launch. Bain & Company, Technology Report 2025, “From Pilots to Payoff: Generative AI in Software Development.”
The review queue
“Our engineers open more pull requests than ever. Our seniors spend the week reviewing, and the queue still grows.”
- What it costs
- Your most expensive hours go to review. Finished work waits, so releases and milestone invoices slip. On fixed-fee work, every extra review round is unpaid.
- What we build
- An agent checks each pull request against its ticket and your team's rules and posts a short summary before a person reviews. It flags oversized pull requests for splitting. A named engineer approves every merge.
- The effect we watch
- How long a pull request waits for its first reviewer, and review rounds per pull request, read from your Git host.
- Industry context
- On teams with high AI adoption, developers merged 98% more pull requests, while pull request review time rose 91%. Faros AI, The AI Productivity Paradox Report 2025, telemetry from more than 10,000 developers on 1,255 teams.
Almost right, then it comes back
“The AI writes it fast. Then someone spends the afternoon making it work, and more of it comes back as bugs.”
- What it costs
- Rework nobody estimated: unpaid on fixed fee, questioned by the client on time and materials. Bugs after release turn into warranty work and incident hours.
- What we build
- Your lead approves acceptance criteria for each ticket. An agent drafts tests from them, an engineer reviews the tests like code, and CI runs them on every change. When a bug comes back, an agent proposes the check that would have caught it, and your lead decides whether it joins CI.
- The effect we watch
- Rework (changes fixed again within two weeks) and bugs found after release.
- Industry context
- The most common frustration with AI tools, cited by 66% of developers: “AI solutions that are almost right, but not quite.” Stack Overflow, 2025 Developer Survey, 31,476 respondents to this question.
Nobody can show what AI changed
“Everyone has an AI tool. Nobody owns whether it changed how we ship.”
- What it costs
- Tool spend and training time with nothing to show in margin or release dates, and no evidence when a client or your board asks what changed.
- What we build
- One owner per workflow, and two numbers: use and the effect your owner chose. We measure from your Git host and tracker for two weeks before any change. Each month an agent drafts the report, your owner checks it, and you decide: continue, stop or extend.
- The effect we watch
- The effect number your owner chose, against its own baseline.
- Industry context
- 90% of technology professionals in DORA's 2025 survey use AI at work. DORA's finding: AI “magnifies the strengths of high-performing organizations and the dysfunctions of struggling ones.” Google Cloud DORA, State of AI-assisted Software Development 2025, 4,867 respondents.
Clients want AI on the invoice
“Clients expect AI speed, and some expect it on the invoice.”
- What it costs
- On time and materials, the same work in fewer hours cuts revenue. On fixed fee, the price gets squeezed at the bid.
- What we build
- Delivery evidence per client. An agent drafts a short “how this month was delivered” section from your tickets, pull requests and hours: what agents drafted, what people checked and which checks passed. Your account lead approves every word the client sees. The pricing decision stays yours, with evidence behind it.
- The effect we watch
- How many client price challenges were answered with evidence, and the outcome, with revenue per hour worked and gross margin per project as context.
- Industry context
- Asked what clients want from them in relation to AI, 45% of agency leaders said faster delivery and 39% said lower fees. Society of Digital Agencies (SoDA), Agency Outlook Study '26, research summary, January 2026, 251 agency leaders surveyed in Q4 2025.
From brief to estimate
“Turning a client brief into an estimate takes our best people days, and the fixed-fee ones still blow up.”
- What it costs
- Unbilled presales hours from your most expensive people, and underestimates that come straight out of fixed-fee margin.
- What we build
- An agent turns the brief or call transcript into a draft backlog with its assumptions and open questions, and shows your closest past projects with their estimate against actual hours. Your lead sets the estimate and owns the number.
- The effect we watch
- Hours from brief to estimate, and estimate against actual on your next closed fixed-fee projects.
- Often the first workflow
Status reports
“My project leads lose Friday to client status reports, and clients still email asking where things are.”
- What it costs
- Lead hours, unpaid on fixed fee and often written off on time and materials.
- What we build
- An agent drafts each report from merged pull requests and ticket changes. Your lead edits and approves it before the client sees it. It's quick to build and easy to see working, so it often comes first.
- The effect we watch
- Lead hours on status each week, and client follow-up questions per report.
Also in scope: client feedback triage, project memory, onboarding and the agents you already run. We test their quality, failure handling and cost, and fix what doesn't hold up. Read the memory guide
Pain 06, drawn before and after
What your free plan draws: the workflow today, and the design we'd test. Each step carries one of the four roles from step 2 of the method below.
Today
Opens Jira, GitHub and Slack for each client
Asks engineers what changed
Waiting on replies
Reply in Slack, when they can
Writes the report by hand
Most of Friday afternoon (from the brief)
Emails follow-up questions
Rework
Answers each follow-up
After
Collects merged pull requests and ticket changes per client
Drafts the report and flags tickets with no update
Edits, checks the flags, approves
Sends it to the client
Logs hours on status and follow-up questions
Key
- Person: a person does it
- AI agent: an agent drafts or prepares it
- Person approves: a person checks and signs it off
- Automatic: runs on its own and is logged
To check: our assumption, not from the brief
- Checks
- Only tickets the client may see
- Internal notes stay out
- A backup approver when a lead is away
- Effect to watch
- Hours on status each week; follow-up questions per report
- Business as usual
- For two weeks the agent drafts while your lead still sends their own report; you compare both before switching
Read the map as text
Today
- Project lead: opens Jira, GitHub and Slack for each client. (from the brief)
- Project lead: asks engineers what changed. Waiting on replies. (to check)
- Engineers: reply in Slack, when they can. (to check)
- Project lead: writes the report by hand. Most of Friday afternoon (from the brief).
- Client: emails follow-up questions. Rework. (from the brief)
- Project lead: answers each follow-up. (to check)
After
- AI agent, automatic: collects merged pull requests and ticket changes per client.
- AI agent: drafts the report and flags tickets with no update.
- Project lead, approves: edits, checks the flags, approves.
- Project lead: sends it to the client.
- AI agent, automatic: logs hours on status and follow-up questions.
How we work
We take responsibility for the result, then hand it to your team.
A named lead and engineers who build. One workflow at a time, in five steps, with your owner deciding at each one. When your people can run it, we step back on a date written into the first agreement.
Weeks 1 to 2
1. Map
We walk through the workflow with the people who do it and read its trail in your tools: tickets, pull requests and chat. We ask what starts it, every step, the decisions, the handoffs, what goes wrong and which tools it uses. You get the map of today.
Dima and our engineer, with your team
Weeks 1 to 2
2. Design
We draw the best version our experience can see, without breaking business as usual. Every step gets a role: a person does it, an AI agent drafts it, a person approves it, or it runs automatically. Anything client-facing or irreversible keeps a person's approval.
Dima · your owner approves the design
Weeks 3 to 4
3. Test
We sit in on real work and interview the people who do it, to check the design against what happens before anything goes into daily delivery. Your owner says go, or we change the design.
Dima and our engineer, with your team
Month 2
4. Build and implement
Our engineers build it inside your systems. It runs next to today's process first; your owner compares the two and decides when to switch.
Our engineer builds · your owner switches
Month 3
5. Monitor
Every month, two numbers: whether it's used, and whether it helps against the baseline. Then we fix it, extend it or map the next workflow. Month three ends with a review: continue, step down a level or hand over.
Dima reports · you decide
Business as usual, by design
- Your tools stay. We build inside the ones you use.
- Nothing switches until your owner says so.
- There's always a way back to today's process, the same day.
- Your owner chooses the first workflow and the effect we watch.
Who leads the work
Dima Levin
Co-founder of Cone Red
150+ production AI projects in the founder's record
Certified Services Partner · Claude Partner Network
Dima leads every engagement himself: priorities, architecture decisions and a regular review with your owner. His monthly hours are printed in your proposal.
Also AI officer at GetDeal.AI.
Production work you can inspect
How we work, in writing
For your CTO and the person who'll own the work: the guides we send with every plan.
Open the guidesThree roles, one result
Dima leads
Priorities, architecture decisions and a regular review with your owner.
Our engineer builds
Implementation and testing in your systems, within the hours of your level.
Your owner runs it
Chooses the first workflow and the effect to watch, approves the design and decides when to switch. Then runs the workflow and checks its output.
Your people run it after us.
The first agreement names the handover date, the readiness conditions and the level we step down to. The three-month minimum applies; staying longer is your choice.
Send one workflow. Get a plan you can act on.
An agent drafts your plan from your answers and public sources, and draws your workflow today and the design we'd test. Dima reviews every page before it reaches you. If the honest answer is to keep what you have, the page says so.
Weekly client status reports, for a 40‑person software agency
Drawn from the answers in the brief. We haven't seen the work. Steps marked “to check” are our assumptions.
- What it costs you today.
- The sample brief: six project leads, most of Friday afternoon each, and clients still email follow-up questions. Nobody has measured it yet; we'd measure it for two weeks before any change.
- The size of the first build.
- About 20 to 30 engineering hours to build and test, which we confirm on the call. Beside it: six leads' Friday afternoons, every week, from the brief.
- The design we'd test.
- An agent collects each client's merged pull requests and ticket changes, drafts the report and flags tickets with no update. Your lead edits, checks the flags, approves and sends it. Steps on your lead's desk: four today, two in this design, untested.
- What stays the same.
- Jira, GitHub and Slack; your lead's approval; what the client receives. For two weeks the agent drafts while your lead still sends their own report, and you compare both before switching.
- Checks before building.
- Only tickets the client may see; internal notes stay out; a backup approver when a lead is away; whether each client's contract allows an AI tool to read its tickets.
- Two numbers to watch.
- Use: reports the agent drafted that your lead approved, and how many they rewrote or wrote by hand. Effect: hours on status each week and follow-up questions per report, against the two weeks before the change.
- Who runs it.
- Your project leads. They learn to adjust the template and check what the agent drafts.
- If you want to take it further.
- 30 minutes with Dima to correct this map, starting with the steps marked “to check”, and agree what's worth scoping.
Sources: the sample brief. Steps 1, 4 and 5 of today's map come from it; steps 2, 3 and 6 are to check. No public sources used.
One page, free, within three business days of a complete brief. No call, no access to your systems. Yours to keep.
Prefer email? Write to [email protected] with the same three answers.
- 3 business days
Your free plan
One written page with your workflow drawn before and after, within three business days of a complete brief.
- 30 minutes
Optional 30-minute call
You and Dima correct the map and agree the work worth scoping.
- 5 business days
A written proposal
Within five business days of the call: the workflows, hours, fee and handover conditions.
You choose whether to sign. Paid work starts after signature.
Hiring an AI lead? Add the job posting. We start the work now and write the handover to your hire into the first agreement.
Choose the capacity your work needs
We recommend Embedded, because the first quarter above needs an engineer building in your systems. Advisory supports a team that builds for itself; Dedicated fits a larger build queue.
The first quarter at Embedded: one workflow mapped, tested, built and measured against its baseline, the next one mapped, and the review that decides your level.
Advisory
AI leadership for a team that does its own building
EUR 5,000 a month
- 8 hours of Dima's time and 12 engineering hours each month
- A weekly review with your owner, decisions written down and one work queue watched.
Recommended
Embedded
AI leadership plus engineering in your systems
EUR 9,000 a month
- 12 hours of Dima's time and 40 engineering hours each month
- Everything in Advisory, plus workflows built and tested in your systems and a monthly review of the two numbers: use and the effect you chose.
Dedicated
A dedicated engineer plus Dima's architecture time
EUR 18,000 a month
- 16 hours of Dima's time and 160 engineering hours each month
- An engineer on your build queue, with Dima responsible for architecture decisions.
Fixed monthly fee. Three-month minimum, then month to month. We invoice on the first of each month. Your proposal prints the hours and scope. Prices are in euros.
Start with the free plan. Choose a level once the work is clear. Get my free plan
FAQ
Our engineers already use Copilot, Cursor or Claude. Why do we need you?
Your engineers keep their tools. We change the workflow around them: which steps an agent drafts, who checks the output, and the two numbers you watch, use and effect.
Why not build it ourselves?
You can build the agent; it's the easy part. The hard part is choosing the step worth changing, getting your people to use it without disrupting delivery, and showing whether it helped. That's the five steps, reported monthly. If you have builders, Advisory fits: we map and design with you, your engineers build, and we review the build and the two numbers.
Will this disrupt live projects?
Your tools stay, and we build inside the ones you use. A new workflow first runs next to today's process, and nothing switches until your owner says so. There's always a way back to today's process, the same day. Anything a client sees keeps the approval of the person who signs it off today.
Our client contracts limit who touches their code.
Then we start on your internal workflows, such as estimates, presales and status reports. We add work on a client's code or tickets only with that client's written permission, and your plan lists that permission under the checks before building.
Why not hire a Head of AI?
You can, and we'll hand over to that person. A senior hire takes time to find and arrives without engineers. We start the work now, and the first agreement names the handover to your hire.
We don't have engineers. Can you still help?
Often a fixed-scope build fits better than a monthly retainer. Your plan will say which.
Three months feels like a long commitment.
You see the proposed workflows, hours and acceptance checks before you sign. After three months the work runs month to month. At the month-three review, if you've used less than 30% of the hours, you move to paying by the hour.
Who owns what you build?
Once it's paid, you own the code we write for you, its configuration, its documentation and your data. Our own tools and methods stay ours, and we license them to you where they're built into your work. Your agreement lists each deliverable.
Can you work with the agents we already run?
Yes. We test their quality, failure handling and cost, and fix what doesn't hold up.
How do we see progress?
Each month you get the work completed, the two numbers for each workflow and how they moved, the hours used and the decisions for next month. Every deliverable has acceptance checks your owner signs.
How do you know it's used?
We log every run as part of the build. Each month you see two numbers per workflow. Use: how often it ran, how much of its output your people approved without rewriting, and how often they went back to the old way. Effect: the number your owner chose, against the two weeks measured before the change. If nobody uses a workflow, we fix it or stop it.
How do scope and support work?
Your proposal names the work, monthly hours and response time during working hours. Work beyond that scope needs your written approval. We keep the price we quote you; a change in scope or level needs your agreement.
What if Dima is busy?
His hours are printed in your proposal and reserved for you each month.
Start with one workflow.
Tell us the workflow, the tools you use and where it gets stuck. Your free plan, with your workflow drawn before and after, arrives within three business days.
Prefer email? Write to [email protected] with the same three answers.
Dima Levin, co-founder, Cone Red