Instacart: Ava, the internal AI assistant — over half of employees monthly, with 20+ minute sessions

• Scaling Productivity with Ava — Instacart's Internal AI Assistant — tech-at-instacart (Zain Adil, Kevin Lei, Ada Cohen), 2023-09-07
• Unlocking Efficiency: How Ava Became Our AI Productivity Partner (43% экономят час+/нед, 60% инженеров, ~70K строк кода/мес) — tech-at-instacart (Dayle Lyng), 2024-01-31

Background

Instacart is the largest US grocery delivery service. The story of its internal AI assistant began not with a strategic program but with a company-wide hackathon in early 2023: working on hackathon projects, an engineering team shipped nearly twice as many features as planned with the help of ChatGPT (specifically GPT-4). The team drew the conclusion quickly: if the model accelerates brainstorming, coding, debugging, and test generation that much, every developer in the company needs access — as fast as possible.

The timing was fortunate too: Instacart worked closely with OpenAI and had early access to GPT-4 with a 32K-token context — through APIs with enterprise guarantees for data privacy, security, and quotas. That removed the main blocker that makes companies ban public chatbots: data does not leak into third-party services on uncontrolled terms. On that foundation the team quickly built and launched Ava — a ChatGPT-style internal assistant, initially for engineers only.

In September 2023 engineers Zain Adil, Kevin Lei, and Ada Cohen described the 'from hackathon project to internal tool' journey on the tech-at-instacart blog with honest adoption metrics, and a second post followed in January 2024 — with data on how Ava took root beyond engineering. Instacart's case became one of the genre-defining public accounts of the 'internal ChatGPT': dozens of companies built similar tools, but few published the engagement numbers and the mechanics that drove them.

Problem

Letting employees use public AI chatbots is risky: code, internal documents, and customer data must not leak into third-party services without enterprise guarantees. But a ban without an alternative has a predictable price — lost productivity and shadow usage: employees will carry work tasks into personal ChatGPT accounts anyway, just without any oversight from security.

So an internal tool is needed — and here comes the second, product half of the problem. Corporate tools are notorious for losing to consumer ones on usability: if the internal assistant is worse than ChatGPT, employees simply won't use it, and the investment becomes a dead portal. The tool had to match ChatGPT on user experience — and add what the public service cannot: integration with the company's workflows.

The third challenge was scaling beyond engineering. For developers the value is obvious from the first prompt: code, debugging, tests. But recruiters, lawyers, marketers, and operations teams have no prompt-writing habit, and a blank text box is a real entry barrier. Rolling the assistant out company-wide is a product problem, not a technical one: you need mechanics that transfer the expertise of power users to everyone else. And all of it had to be measured — not 'registration counts', which internal PR can easily inflate, but real engagement and retention.

Solution

Ava is a ChatGPT-style web assistant on GPT-4 (including the 32K-context model) and GPT-3.5. For the first launch — engineers only — the team prioritized features with instant developer value: convenient keyboard shortcuts, single-click code copying, and automatic upgrades between GPT-4 models as the conversation context grew — the user never has to think about which model they need. With the largest context model, developers worked with whole code files: creating, debugging, reviewing — and summarizing documents.

Adoption outpaced expectations, and post-launch interviews revealed demand beyond engineering — from Ops and Recruiting to Brand Marketing and HR. For non-technical users the team lowered the 'blank box' barrier with well-crafted prompt templates, added full-text conversation search and conversation sharing. The key sharing detail: conversation links unfurl into previews in Slack so colleagues see the content before clicking — the team notes directly that this significantly aided adoption and product awareness.

The next step was the Prompt Exchange — an internal prompt marketplace. The logic is instructive: a large share of conversations started from templates, but Ava's small team physically lacked the domain expertise to write good templates for every department. Instead, publishing and starring prompts was handed to employees themselves — best practices began spreading horizontally, with no central team involvement.

The fourth step was going where employees already work: a Slack app. Mentioning @Ava in a thread produces summaries and answers right in the conversation — 'like asking a colleague'. By then nearly a third of the organization was already using Ava monthly, and the Slack channel removed the last friction — having to open a separate page and paste context by hand.

The plans described in the post: retrieval over Instacart's internal knowledge and code execution (the team calls data the 'Achilles' heel' of LLMs), plus opening Ava's APIs to every team in the company so they can build their own AI tools on top of the assistant.

Result

As of the September 2023 post, over half of Instacart employees used Ava monthly and more than 900 weekly; sessions ran 20+ minutes, and users 'were producing and copying a significant amount of code'. The January 2024 follow-up added specifics: 43% of the company saves more than an hour a week with Ava; 60% of engineers generate around 70,000 lines of code with it monthly; the Slack plugin is invoked over 5,000 times a month and summarizes more than 200 threads and channels. All figures are Instacart's internal data without external audit; note also that 'an hour saved per week' is employees' self-assessment from surveys, not a time-tracking measurement.

In our view, the Ava case is valuable above all as a textbook of adoption product mechanics. Internal tools are usually rolled out by decree; Instacart instead walked the classic consumer path: start with an audience that gets instant value (engineers), lower the entry barrier (templates), viral loops (conversation sharing with Slack previews), user-generated content (the Prompt Exchange), and going to where the user already lives (the Slack bot). Every mechanic here transfers to any company — and judging by the retention numbers, together they work better than any corporate mandate.

Our second observation concerns metrics: Instacart reported engagement, not registrations — weekly audience, session length, volume of generated code, Slack invocation frequency. For internal AI tools that is rare discipline: 'we gave everyone access' and 'half the company actually uses it every month' are fundamentally different statements. Finally, the economics of the bet is telling: a hackathon project received a product team and enterprise security guarantees within months — a speed of legalization that likely explains why shadow AI use never had time to take root.

Lessons learned

1. An internal ChatGPT with enterprise guarantees is the fastest way to legalize AI at a company: employees get the tool, security keeps control of the data.
2. Product mechanics drive adoption, not mandates: the Prompt Exchange and conversation sharing with Slack previews turn colleagues' best prompts into a viral distribution channel.
3. A blank text box is a barrier for non-technical users: ready-made prompt templates opened Ava to Ops, Recruiting, Marketing, and HR.
4. Go where the user already works: the @Ava Slack bot removed the separate-page friction — 5,000+ monthly invocations confirm it.
5. Measure engagement, not signups: 20+ minute sessions, 900+ weekly users, and 70,000 lines of code a month signal real utility rather than curiosity.
6. The central team need not be the expert in every domain: hand template and prompt creation to the users themselves — as the Prompt Exchange did.
7. A hackathon is a legitimate source of corporate AI infrastructure — as long as the project quickly gets a product team and security guarantees.

Frequently asked questions