Build it yourself? The real 101-task checklist

01 Connect your tools — and keep them connected A demo connects one tool, once. A product connects 143+ and keeps every one syncing through every outage and silent API change. 14 tasks

Build secure connections to 20+ tools (Google Workspace, Microsoft 365, Slack, GitHub, Jira, Zendesk, Zoom, Linear, Asana, Pendo, Snowflake, phone systems…), each with its own login, quirks, and limits

Add 100+ more CRM, HR, and ticketing systems through unified-API plumbing

Build custom connectors for tools with no off-the-shelf path (Snowflake key-pair auth, Salesforce Service Cloud, and others)

Refresh every tool's access automatically before it expires — with a lock so two jobs don't refresh at once and collide

Handle each provider's different refresh rules (Microsoft doesn't work like standard logins)

Detect when a customer revokes access and notify your team before the customer notices

Sync only what changed since last time, in each provider's required timestamp format

Run separate, orchestrated jobs to backfill years of history without melting the live sync

Mix real-time push (e.g. Gmail notifications) with scheduled polling for everything else

Survive providers throttling you — one connector retries up to 1,000 times with randomized backoff, because that's what the API actually demands

Give each tool its own job queue so one slow API can't starve the rest

Skip and log permission-denied items instead of failing the whole run

Normalize wildly different data shapes from every tool into one consistent format

Carry dozens of workarounds in code for documented-but-broken API behavior

…then maintain all of it as each of those 143 tools changes on its own schedule, with no warning.

02 Figure out who's actually who The same person shows up as six different IDs across your tools. Get this wrong and every number silently double-counts or misses people. 7 tasks

Pull every email alias from Google and Microsoft directories and resolve them to one real person

De-duplicate people who appear in multiple systems the same way every time, so re-runs don't reshuffle identities

Map Slack IDs to the right person, including fuzzy name matching

Infer when someone joined or left from their activity, when HR data is missing or wrong

Stop counting departed employees in current numbers

Detect meeting rooms and shared mailboxes pretending to be people, and exclude them

Build manager/report org-chart views — including custom loaders for non-standard org data

03 Turn raw activity into numbers that mean something "Meeting time" and "focus time" sound simple. They hide a dozen ways to be wrong that are obvious to your customer and invisible to you — until after launch. 13 tasks

04 Build the assembly line that produces every number None of those numbers come from the AI. They come from a 900-step pipeline that has to run in exactly the right order, every day, for every customer. 7 tasks

439 data models across staging, intermediate, metric, and final layers

~500 transformations that must run in a strict order (threads before tenure before meeting-eligibility before metrics)

69 custom warehouse functions for business-hours math, timezones, and domain logic

100+ files of encoded reference knowledge

Run it two ways: a full-history rebuild per customer, or a rolling daily update

Tune it for many customers at once so queries stay fast

Sync data both directions between the app database and the warehouse, every run

05 Catch problems before your customers do A dashboard that's quietly stale or wrong is worse than no dashboard. This is the layer that keeps the numbers honest while you sleep. 7 tasks

Track sync status and freshness for every tool, per customer

Check freshness deadlines every 30 minutes and page the team on a breach

Test for upstream providers silently changing their data format

Detect duplicate records by customer and tool

Audit permissions daily, catch lost access, and auto-rebuild when it's restored

Requeue dead jobs automatically; flag any backed-up queue

Route sync alerts so the team sees signal, not noise

06 Don't leak everyone's private data To produce these numbers you're handling the whole company's messages, calendars, and HR data. One mistake here is a breach, not a bug. 11 tasks

07 Make it work differently for every team Every customer measures themselves differently. The platform has to bend to each one without forking the code. 5 tasks

Per-customer settings for KPIs, digests, business hours, and domains

80+ feature flags scoped per customer

Custom data imports to add customer-specific attributes

Isolated demo environments so sales demos never touch real data

Capacity and plan tracking with history

08 Run the AI itself — the part everyone thinks is the whole thing Even the easy 5% is real engineering once it has to run reliably, at a sane cost, for everyone at once. 7 tasks

Break large context into capped chunks, summarize, then combine the results

Rate-limit per model on both requests and tokens

Route across multiple model tiers, switchable on the fly

Retry failed model calls with backoff and sane caps

Profile latency at each stage

Stream answers over pre-computed, identity-resolved data — not raw exports pasted into a prompt

Let customers customize prompts; generate coaching plans

09 Push the right thing at the right time "Ask the AI a question" is pull. A system people actually trust pushes the right alert at the right moment — without spamming anyone. 7 tasks

A rule engine for conditions, actions, and alerts

De-duplicate recurring alerts so the same issue doesn't fire twice

Score urgency, with preview-before-activate and per-account snoozing

Schedule digests honoring each person's frequency, day, hour, and timezone

Daily and weekly alert pipelines; a daily briefing with its own health check

Make scheduled reports post exactly once, never twice

Deliver across email and Slack, with tracking

10 Build the actual product people open After all of that, you still have nothing anyone can use. This is the front-end that turns the data into a product. 10 tasks

11 Keep the lights on, 24/7 It's not a project that ships and ends. It's a service with an SLA that someone is on call for — forever. 8 tasks

Define ~40 cloud resources per environment as code (servers, networking, load balancing, firewall, certs, alerts)

Patch operating systems monthly, draining work gracefully first

Full observability: traces, metrics, logs, and error tracking with user context

Health endpoints wired straight to alerting

Enterprise login: SSO across SAML, OIDC, Okta, and Azure, with domain matching

An internal toolset for deploys, scrubbed data copies, consoles, secrets, and health checks

Reproducible environments and separate deploy tracks for front and back end

364 database migrations of accumulated change, plus a real test suite and CI

12 Catch wrong numbers before your customers do A build that fails loudly is the good case — the dangerous one looks fine and is quietly wrong, and you find out when it costs you a customer, a hire, or a board number. 5 tasks

Detect when a metric silently returns nothing instead of erroring (a blanked number reads as “—”, not as a failure)

Reconcile the same metric computed in two places so they never silently disagree

Alert when a data sync reports healthy while updated rows sit behind a stuck checkpoint

Validate that exclusion rules (vacations, weekends, internal time) are applied identically across every code path

Version every metric so any past number can be reproduced and audited

So you want to build it yourself?

That's 96 distinct jobs across 11 areas — and that's just the boiled-down version.

How many could your team honestly check off today? And who maintains the rest?