# Rollout The code is done. This is the part that decides whether it earns anything. --- ## Pilot on your own support desk first Not a hospital. Your support team. Three reasons it is the right first deployment: 1. **The problem is already documented there.** Replies going out to clients in English that would not read well if it were ever produced as evidence. Sahi's Hinglish rewrite and the legally-cautious rewrite prompt were written for exactly that. 2. **Zero sales risk.** If the selectors break on Gmail, or the underlines land two characters off, you find out from your own staff and not from a hospital's medical director. 3. **It produces the pitch.** Two weeks of Insights on your own team gives you a real number — "our support desk was writing *revert back* eleven times a week" — and a screenshot with nothing confidential in it. That slide sells better than any feature list. ### Two weeks, concretely **Before day 1** - Stand up the server. Run the Health page until every row is green. - Change the seeded password. Set `allowed_origins` to your real origins. - Turn Insights on for your own org. It is off by default. - Issue one key per person, named properly — *Support — Ramesh*, not *key 4*. The by-writer report is only useful if the labels are. **Week 1 — quiet running** - Five people, not the whole team. Pick two who write well and three who struggle; you need both to tell signal from nagging. - Tell them plainly what it sends and what it keeps. Point them at the privacy policy. Somebody will ask, and the answer being ready is the difference between adoption and suspicion. - Do not announce a target. You are measuring the baseline. **Week 2 — tune** - Open Insights. Any rule sitting under 40% acceptance is arguing with people who are right. Switch it off. `\bu\b → you` will probably be the first to go. - Add every drug name, doctor name and product name that got flagged to the dictionary. This is the single highest-value hour anyone will spend on the product. - Ask the five what annoyed them. Underlines that are wrong cost more trust than underlines that are missing. **What to measure** | Signal | Where | What good looks like | |---|---|---| | Acceptance rate | Insights | Above 70% overall. Below that the rules need pruning, not the team. | | Checks per person per day | Insights, by writer | Rising through week 1. Flat at zero means somebody turned it off and did not say. | | Rewrites used | Usage, endpoint `rewrite` | Any use at all. This is the feature that justifies the AI cost. | | Complaints about wrong flags | Ask them | Should fall to near zero after week 2's tuning. | **Do not measure people.** The by-writer table is a coaching list. The moment it becomes a scoreboard, staff stop clicking Accept and start avoiding the fields, and you lose both the data and the benefit. --- ## Then one hospital Pick the client who already asked for spell check in HIS. They raised the request; you are answering it, not selling them something. - Deploy the **embed widget** on one screen — discharge summary is the obvious one. Not the whole product, one screen. - Insights **off** for them unless they ask. Do not create a data question you did not need to have. - If they are on-prem and cautious, use the Docker stack with the AI switch off. Rule pack and LanguageTool only, nothing leaves the building, and you can put that in the contract. --- ## Product sheet One page. Adapt the bracketed parts. --- ### Sahi — writing assistance for hospital teams **The problem.** Discharge summaries, client emails and WhatsApp replies go out with spelling and grammar errors. Some are embarrassing. Some end up in a medical record, or in a dispute, where the wording matters. Generic tools do not help much: they flag every drug name and know nothing about how English is actually written in an Indian hospital. **What Sahi does.** Checks writing where it happens — the HIS screen, Gmail, WhatsApp Web — and offers the correction. One click to take it. **What makes it different** - Knows the vocabulary. Drug names, clinical abbreviations and doctor names stop being underlined; your own terms are added in minutes. - Knows the habits. *Revert back*, *do the needful*, *informations*, *kindly intimate* — and turns *pls / plz / thanx* into something you can send a client. - Turns Hinglish into professional English in one click. - Rewrites for tone without inventing content: it will not add a promise, a deadline, an apology or an admission that was not already written. **Where it runs.** Inside Caresoft HIS text fields, and in the browser for email and WhatsApp. Same engine, one place to manage it. **Where your data goes.** Your server. Sahi is self-hosted — on your hardware if you want it there. With the AI option switched off it runs entirely on your own infrastructure and no text leaves the building. Message text is never stored beyond a short cache you can switch off. **What administrators see.** Which terms your team keeps getting wrong, whether the corrections are being accepted, and per-person usage. Reports count phrases, never messages — anything longer than six words, or containing a number, is discarded before it is stored. **Deployment.** Managed install across every machine through standard Chrome policy. Nothing for staff to configure. *[Contact block]* --- ## Pricing structures Three shapes that fit this product. Each has a real failure mode; none is obviously right, and the choice depends on how you already sell HIS. **Per named user, per month.** Simplest to explain, matches the per-key model already in the product, scales with value. *Failure mode:* hospitals count seats down to the bone and you end up with eight licences for a forty-person department, which kills the Insights data that makes the product sticky. **Per hospital, per year, banded by beds.** Matches how you already price HIS, so it drops into an existing renewal conversation with no new negotiation. *Failure mode:* the small nursing home pays almost nothing and still generates support load; and you carry the AI cost without a lever if one site is heavy. The per-org monthly cap in Tenants exists for exactly this. **Bundled free with HIS AMC, sold separately elsewhere.** Turns Sahi into a reason not to churn, and the AMC hike conversation gets easier when something new arrives with it. *Failure mode:* you never learn what it is worth, and it becomes very hard to charge for later. Two things that are true whichever you pick: - **On-prem, AI-off should be cheaper, not free.** It costs you nothing per check, but it is the version a cautious hospital wants, and free trains them to treat it as a feature of HIS rather than a product. - **The AI tier has a variable cost you do not control.** The rule pack and LanguageTool are free forever; every model call is not. Set the monthly cap on every tenant at signup, not after the first surprise. I am not a financial adviser and these are structures rather than a recommendation — the numbers depend on your AMC base and cost per check, which you can read off the Usage page after the pilot.