Context is king: How teiō uses AI

We started this newsletter to share my thoughts after speaking with 100+ companies over the last six months on what they're doing with AI. This newsletter is not currently public and only meant to share our mental models with our leads, customers and partners.

Today's topic is: context is king, and why we don't use any external tools at teiō.

Before I show you how teiō uses AI internally, let's chat about how we think about this space. Everyone has access to the same models now, so the one big advantage left is context. From first principles there are only two ways to get it:

1. Your data lives in many different systems. You centralize it in a data warehouse, clean it, structure it, and model it. All of that foundational data work is what gives your tools the right context. For 99% of enterprise companies this will be the right approach.
2. You minimize external software and build everything purposefully in a single system. We would still recommend pulling it into a warehouse and keeping it clean and structured, but that step becomes a lot easier. Same outcome, with more personalization and speed, and less cost and maintenance.

Our thesis is the data warehouse is the best avenue to win the context race. You can build agentic applications other ways, but the warehouse is the winning paradigm for the most complete, most reliable context.

Being an AI-first consultancy, we had the opportunity to take the second approach. We built our own internal operating system, similar to Glass at Ramp, and 100% of our team runs on it every day. teiō does not use any external software apart from Roam, which is our virtual office and company-wide chat.

The platform contains:

• CRM
• Proposal generator
• Project management and delivery
• Customer-facing portal
• HRMS
• Gamification, performance and engagement
• Financial hub
• Context hub
• Ask AI
• Notetaker

Here's a quick tour. Everything below is from teiō play, a demo instance of our platform with sanitized synthetic data. The players run at a steady pace: pause, step frame by frame, or change the speed.

CRM

⏸‹›1 / 10

0.5×1×1.5×2×⛶

Every customer in one place. Open a company and you get the points of contact, billing, linked proposals and projects, and a full conversation timeline: every email, meeting, and channel message with that customer.

Proposals

⏸‹›1 / 16

0.5×1×1.5×2×⛶

We've automated 90% of proposal building. AI breaks the work down using our methodology, looks at past tasks and how long they took, gauges relative complexity, and uses statistical modeling to assign value. AI drafts the proposal and lays out the payment options. Customers view the full proposal with pricing and accept online. We then track sentiment, close probability, and estimated start date, which we use to plan resources.

Delivery

⏸‹›1 / 9

0.5×1×1.5×2×⛶

Every project broken into phases, every task with an engineer and story points, so you can see exactly what's done, what's in progress, and what's left. Scope changes come in as change requests the customer approves right in the portal.

Gamification, performance and engagement (AI enablement)

⏸‹›1 / 12

0.5×1×1.5×2×⛶

This is the end state for AI enablement: you can't enable something you can't measure. Payouts depend on performance and how much work engineers actually ship. Everyone starts on a standard base salary, and our top engineer makes 2 to 3x what a good engineer makes. We also measure engagement: activity on the platform, how much AI and Claude Code people are using, time in Roam (our virtual office), all-hands participation, role channels, and surveys. It all rolls into an engagement score out of 10, and the wall of fame crowns each month's champions.

HRMS

⏸‹›1 / 14

0.5×1×1.5×2×⛶

The directory with an org chart view, profiles, roles and exactly what each one can access. HR gets leave tools, onboarding, offboarding, anonymous surveys, all-hands tools and per-person all-hands attendance scoring.

Finance

⏸‹›1 / 5

0.5×1×1.5×2×⛶

Accounting and banking. Live cash across every account, card, and currency, plus monthly customer statements generated straight from delivery data. More to come here: projections, AP and AR optimization.

Ask AI

⏸‹›1 / 12

0.5×1×1.5×2×⛶

Ask AI is the glue which brings everything together. Here it's answering "How do we find $500k of new revenue?" Watch the reasoning steps. It pulls the lead pipeline and customer roster, spins up parallel research across accounts, reads each briefing, understands sentiment and comes back with a ranked plan. When I ran this on our production instance it goes 67 reasoning steps deep before answering.

Notetaker

⏸‹›1 / 8

0.5×1×1.5×2×⛶

We built our own notetaker so the full context of every call is available to the right folks out of the box. It reads the calendar and joins calls automatically. Customer calls are always recorded, the rest is a toggle. Every call is filed against the right company with the summary, the transcript, who was on it, and who can see it. And since the call lives next to everything else, you can ask AI about it right from the call page.

What it costs us to run

Contrary to market sentiment, we are not spending $10k/month on tokens. The platform costs us:

• Hosting: $26/month
• AI tokens: $110/month
• External databases, transcripts, other tools: $20 to 30/month

Under $200 a month, and it replaces our entire software stack. Separately, the full company has Claude teams, and users are moved to higher tiers when we see their engagement and usage of Claude Code go up. When you purposely build something it's incredibly efficient and an order of magnitude cheaper than rolling out SaaS off the shelf, before you add up the external cost of maintaining those tools. Note: our results are not typical. Your results will vary based on your data volumes, size and usage. None of this is indicative of what your costs would be.

We're opening this up

We're now at a point where we have started building 100% custom operating systems for companies. We've done pilot projects and the results speak for themselves. Here's how it works:

• A 100% custom operating system built for you
• Initial version live in 4 to 6 weeks, complete build in 3 months, and month 4 is focused on training and transfer. We have developed agents that pull from the teiō codebase to build your own OS. That's how a rollout that took us 6 to 8 months compresses into weeks.
• You own 100% of the code. Zero licensing fees.
• Built for a handoff, not for us to own. We teach your engineers agentic software engineering so your internal team can maintain and extend it after handoff.
• Pricing is based on output: we charge based on the modules you want us to build. We don't charge SaaS, for time, or for engineers.

Note these times can vary depending on the number of modules and the complexity of your use cases. The timeline above is what we expect for a six-module system. The teiō platform currently is a 10-module system with the modules listed above.

FAQ

What exactly do you get?

Your own custom operating system. Modules can cover CRM, GTM, marketing, finance, delivery, HRMS, operations, research. We scope the modules around your requirements.

How do we price?

Implementation fees are based on what you want us to build. We charge for output, not time and materials, and there are no SaaS fees. Packages range from monthly pay-as-you-go to pre-purchasing credits upfront.

How long does it take?

A lot depends on what you choose and the complexity, but you can expect a first version up and running within 4 to 6 weeks and a production-grade system within 3 months, followed by handover training.

Can you build things other than operating systems?

Yes. We now support the full spectrum of agentic software engineering, using agents to build software.

What about security?

On hosting, we deploy a recommended stack focused on multiple redundancy: databases and blob stores with retention policies and daily backups. On platform security, we run regular audits during the build phase to look for vulnerabilities and threats. On governance, everything is built with role-based access control, so you define exactly who can access what across the platform.

Which foundation models do you support?

We build using Claude, either through our account or your own Claude subscription. Inside the platform you can plug in any API. We use our direct Claude API key, plus the Kilo gateway to support models outside Claude when required. Here's the model settings screen, where you pick which model runs each AI task.

What do you use to build?

gstack on top of Claude Code for agentic software development. We built and customized gstack to work exactly on our methodology and process.

What about vibe coding? What are the risks?

Vibe coding is a completely different animal from software engineering. We use pre-trained skills for particular personas (CEO, engineering manager, release quality, security, retro) to mimic the complete software engineering process. Writing code in the traditional sense will live on for mission-critical applications. For the large majority of companies building internal operating systems, agentic software engineering with best practices, processes, and guardrails is the future.

What prevents us from doing this in-house?

Absolutely nothing. We've spent the last 6 to 8 months going deep here, building our own operating system and testing it on teiō, where 100% of the company uses it today. So we understand the quirks of how these things work, which assumptions are load-bearing, and the difference between building something that works for a demo and building it for production. Our advantage is that we're building these operating systems using what we have learnt. Our agents have full access to teiō's code base and routinely pull entire chunks of code as a starting ground. That's how we bring something to market in 4 to 6 weeks instead of the 6 to 8 months it would probably take yourself.

What about maintenance and updates?

We don't believe in SaaS long term. So our default is build, train, transfer: we teach your team agentic software engineering and you own future development. That said, some companies won't want to build that skill set, and you'll be able to buy support and maintenance packages. As we do more of these, we will have a larger collection of pre-built modules to pull from and updates to existing ones. All of it stays output-based: we scope exactly what's needed and charge for that.

Can you advise us instead of building?

We do offer fractional chief AI officer style engagements where we advise you on how to do this. However, what we've found works best is still build, train, transfer: have us come in and build a few modules so you can see how it's set up, then we train your team to take over.

What's the long-term strategy?

Our job is to be at the forefront of two areas: the data/AI axis, and agentic software engineering. We invest in R&D, build expertise, agents, and processes, then sell the output while continuing to invest in R&D. We can't uplevel you beyond our internal sophistication level. The proof this works is that we use it to run our own business.