AI agent development San Francisco companies can trust with one job and one number.

I'm Yiannis Papastergiou, an independent AI consultant based in Thessaloniki, Greece, working with companies of roughly €1M+ revenue. For AI agent development in San Francisco I build one agent for one job, with written limits, inside the tools your team already uses. I have no San Francisco office, team or clients; the work is remote.

AI agent development · San Francisco · Remote from Greece

About 3–5 minutes, mostly one-tap answers, then your name and work email. Your result on screen, with a 7-day plan. No call required.

Take the free AI Growth GPS

See the AI P.R.O. Method

What is AI agent development for a San Francisco company?

AI agent development for a San Francisco company means building one agent that does one defined job inside your existing tools, with written limits on what it may not do, and one number agreed in advance. I build it remotely from Greece in six weeks as part of Run, and the last payment waits until the number is met.

The short answer

If San Francisco is full of agent builders, why do agent projects still stall?

Finding someone who can build an agent is rarely the hard part in San Francisco. Choosing the job, setting its limits and keeping it useful after the builder leaves is harder.

Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. SD Times, reporting the prediction, also noted Gartner's warning about agent washing: vendors relabelling assistants, RPA and chatbots as agentic. It is a forecast from 2025, not a measurement, but the three reasons are the ones worth designing against from day one.

An agent without a named job, a number and written limits is a demo with a bill.

Each of the three reasons becomes a question on the first page of the plan.

The problem

Which agents pay first for a San Francisco software or services company?

Illustrations of the kind of job, not client cases. In each one a person approves anything that leaves the building.

Reads an incoming ticket, finds the customer's plan and recent activity in your tools and drafts a reply for a person to approve. The number: time to first reply.

Support triage

Before a sales or partner call, assembles what the company does, what changed recently and which of your accounts overlap. The number: preparation minutes per call.

Pre-call research

Drafts answers from your own approved policies and flags anything it cannot support. The number: days to return a questionnaire.

Customer security questionnaires

Answers 'how do we do this here?' from your own documents with the source shown, so new hires stop interrupting senior people. The number: time to find the right procedure.

An internal knowledge assistant

One agent, one job. The next one comes in the next 90 days.

Where agents pay first

What should an AI agent at a San Francisco company never be allowed to do?

Every build starts with a short written list of limits, agreed before anything is built.

If the document is not in the approved set, the agent says so and hands over to a person.

Answer from outside approved sources

Drafts go to a person. Autonomy is earned one step at a time, against the measured number.

Send anything to a customer unreviewed

Reading is the default. Writing to your CRM, billing or production data needs its own explicit decision and an activity log.

Change a record in a system of record

Hiring, credit, pricing for an individual and similar calls stay with people, partly because of the California rules below.

Make a significant decision about a person

What it may not do

Which California rules should an AI agent be checked against?

I am not a lawyer, so treat this as a question for yours.

The California Privacy Protection Agency's regulations, approved in September 2025, took effect on 1 January 2026. Businesses that use automated decisionmaking technology to make significant decisions must comply beginning 1 January 2027, and businesses subject to the risk-assessment rules had to begin complying on 1 January 2026, with an attestation due to the agency by 1 April 2028.

An agent that only drafts for a person to approve sits differently from one that decides. That is one reason the first build keeps people in the loop. If your product embeds a model for public use, California's AB 2013, effective 1 January 2026, adds a separate training-data disclosure question, and Crowell & Moring noted in December 2025 that it sets no threshold for substantial modification.

What applies

A San Francisco AI agent development company, a platform agent or a remote independent?

Option

Fits when

Watch for

Plan, from €1,500, can come first if you are not sure which job to build.

San Francisco agent development company

You want a team on site and a large build

Check how the number is agreed before the scope

A platform's built-in agent

The job already lives inside one product

It cannot cross your other tools, and limits are the vendor's

Remote independent, me

One job, one number, your existing tools

I am remote from Greece, with no San Francisco presence

Three ways to get an agent built, and what each asks of you.

Options compared

How does remote AI agent development work for a San Francisco company?

San Francisco is 10 hours behind Greece at the moment, so the two meetings sit in your morning, my early evening. The rest runs asynchronously, so the time difference rarely matters. I do not travel to San Francisco for this work.

Access in place. The clock starts here, not when you pay.

Kickoff

The first of two meetings. You name one owner on your side; your admin adds my setup once.

A 3-minute video and the scoreboard every week. No status meetings.

Weeks 1–6 · The build

I fix the process first, then build, inside the AI tools your team already uses.

A guide and a recorded walkthrough.

Handover

The second meeting, recorded so new people can learn it later.

The last payment, once the number is met.

Day 90

The number, measured against today's baseline.

How it runs, remotely

Who is the AI agent developer behind this page?

You deal with the person who does the work.

At Companies House, the UK's official business registry, I worked on the systems behind millions of searches and filings every year, and was security and quality champion.

Software that must not break

As Product Owner at Natech I led its anti-money-laundering software from desktop to a modern web platform.

Regulated fintech

For Bronzed Studio I automated the Instagram booking flow, and the owner got back at least 10 hours a week.

A small example

I have no San Francisco clients yet, so judge my thinking on the free AI Growth GPS before you spend anything.

Who builds it

Which services sit around an AI agent?

Regulated business? AI Governance & EU AI Act Compliance can be added at any step.

An agent is part of Run. Plan decides which job to build, and Own keeps it running.

Where you stand, and a verdict.

AI Readiness Assessment

What is worth building, in order.

AI Strategy & Roadmap

Plan

A system that does the job.

AI Agents & Custom AI Solutions

Repetitive workflows, measured in hours back.

Automation

Run

An owner for AI, embedded.

Fractional Chief AI Officer

Own

For your business

Questions San Francisco companies ask about AI agent development

A fixed investment, quoted in your Plan, and the last payment waits until the agreed number is met. The only figure I print on this page is Plan, from €1,500, with a seven-day refund. Everything built stays yours.

What is the investment for AI agent development in San Francisco?

Built in six weeks and proven in 90 days. The clock starts when access is in place, not when you pay. You get a short video and the scoreboard every week, with two meetings in total: the kickoff and the handover.

How long does an AI agent take to build?

No. I am based in Thessaloniki, Greece, and all work for San Francisco companies is remote from Greece. I have no San Francisco office, no local team and no San Francisco clients yet. If you want a team in your office every day, a local firm is the better fit.

Do you have a San Francisco office or San Francisco clients?

Whatever your team already uses, so it is delivered into those tools rather than as a new app to learn. Each person signs in to their own accounts, guided step by step. I never ask for passwords.

Which tools will the agent work inside?

Then the last payment is not invoiced. Every build has one number, agreed in writing before we start. We set the target in Plan and Day 90 measures it. If a job cannot show a return on paper, I will tell you not to build it.

What if the agent does not hit the number?

I name the job, the limits and the number first, and I build in your tools rather than reselling a platform. Gartner has warned that vendors relabel assistants and chatbots as agentic. Ask any builder, me included, what the agent may not do.

How do you avoid agent washing and vendor hype?

They shape the first version. The CPPA's rules on automated decisionmaking apply from 1 January 2027 to businesses that use it for significant decisions, so the first agent drafts for a person to approve instead of deciding. I am not a lawyer; your counsel answers the legal question.

Do California rules change what you build?

Yes. Start with the free AI Growth GPS, which takes 3–5 minutes. If you want to decide what to build, Plan is from €1,500. Each step ends with something you keep, and none is a condition of the next.

Can I start without commissioning a build?

Questions

Not sure which agent your San Francisco company should build first? Start with a free 3–5 minute conversation.

About 3–5 minutes, mostly one-tap answers, then your name and work email. Your result on screen, with a 7-day plan. No call required.

Take the free AI Growth GPS

A guided conversation with my AI Growth GPS. Tap or type your answers. It is free so you get something useful before we ever speak.

Discover · The free AI Growth GPS

Not another demo. A number that moved.

Start with the free AI Growth GPS. No call required.

About 3–5 minutes, mostly one-tap answers, then your name and work email. Your result on screen, with a 7-day plan. No call required.

Take the free AI Growth GPS

See the AI P.R.O. Method