AI systems in the software you already pay for

AI systems are the service most people mean when they say implementation or custom AI builds. Nidrosoft sells them after a map exists. That map can come from our audit or from a packet you already trust. Scope defines what the machine does, what a person still approves, and what you will measure. Build puts the system in your tenant. Own means the work lived in your tools from the first accepted commit. A new app with no existing system of record belongs on Product.

The method words still apply. We embed with the operators even during build, because the Friday workaround changes. We re-diagnose if the hours were wrong. We refuse a scope that writes money, contracts, or customer messages without a human. Those rules are how clinics, lots, and landlords stay out of preventable errors.

Public systems show the texture of a finished thing. Omnira Dental for practice operations. Octana for a finished video from a message. Eviction Wizard from notice to writ. Guidera on iOS. TokenTra for spend, with teams cutting 10 to 30 percent of wasted tokens. Protectron for EU AI Act artifacts. Nerlude for infrastructure control. Nexuvo for dealer inventory, 3x more cars and 80 percent less manual outreach. NanoBrief for briefs in under two minutes. Clear Audio AI and MixGuidera for sound. Your build will be named after your process, not after those brands, unless you license a product.

A useful scope sentence looks like this: overnight, the desk contacts private sellers who match the buy box, a person approves any offer, and the CRM shows the thread. That is a Nexuvo-shaped result. Another: the practice receives a complete intake before the chair, a person confirms clinical exceptions, and no-shows drop by a number you already track. That is an Omnira-shaped result. We will not start a build on “add AI to the CRM.”

Scope also names the non-goals. No homepage chatbot if the hours are in the back office. No voice agent if the phones are already fine and the pain is reconciliation. No multi-department rollout in month one. Non-goals protect the operator from a steering committee that keeps adding wishes.

How-we-work describes a common shape: audit in weeks one and two, plan in the following window, implement and train in a later stretch often listed as weeks five through ten. Your dates will move with access, vendor APIs, and the weight of the first process. We would rather rewrite the calendar than ship a sandbox and call it production.

Price follows that scope. This page will not invent a retainer or a currency figure. Book the audit. If you already have a map, bring it and we will price the first outcome.

Scope reviews happen in writing. A verbal yes in a hallway is how extra agents appear. If a sponsor wants a second process in the same calendar, we reopen Diagnose for that process. Sometimes the second process is cheaper because the connector already exists. Sometimes it is a new product. The packet should say which.

We will also write what happens when the model is wrong. A retry. A queue. A human. Silence is not a design. Eviction files, dental exceptions, and dealer offers all fail in public if the fallback is hope. Your runbook gets the fallback in the same week as the happy path, not as a later phase that never arrives.

If your internal platform team already owns agents, we can build a skill they will maintain. If they do not exist, we still leave files they can hire against. Either way the Tuesday operator remains a named person, not a distribution list.

We build in your cloud or your SaaS tenant when the work allows it. Azure, AWS, Google Cloud, Cloudflare, private infrastructure, multiple tenants. EU hosting and EU data residency when required. Approved access. Audit trails. If counsel needs a region, write it in the first week. San Diego is where Cyriac lives, not where your data is required to sit.

Connectors are part of the build. A model that cannot see the dealer board or the practice schedule is a chat window with extra steps. We will say no to a scope that assumes a human will paste every record into a prompt forever. That is a seats problem, and seats are cheaper.

Keys live in your vault. Nerlude exists because founders lose credentials in Slack. We will not recreate that mess. Logs stay where you can keep them. If a vendor’s agent platform hides prompts and traces, ask whether you still own the system after they raise prices.

Staging matters. We will ask for a copy of the CRM, a sample of the inbox, or a time-boxed production role with logging. Building only in a designer’s head produces a demo that dies on the first real record. If security needs a week to grant the copy, the calendar waits. That wait is cheaper than an incident.

We will document the write permissions in a table the operator can read. Which object. Which field. Who approves. What is logged. If that table cannot be explained to a new hire, the build is not done. Fancy architecture diagrams that only we understand fail the Own test.

Model choice is a means. Claude, ChatGPT, Gemini, or a hosted model on your cloud can all appear. We pick for the task and the policy, not for a sponsorship. TokenTra-shaped visibility belongs in any build that will spend real money on tokens. If you already waste 10 to 30 percent, fix that in parallel or the new agent will hide in the same invoice.

Someone on your side pairs through the build. They do not need to be a research scientist. They need to be the person who will operate the workflow in month three. We use that pairing to find who your AI players are, the phrase from the method page. Those people become the trainers for the next team if you later buy transformation.

Cyriac Zeh is the studio side of the pair. Nidrosoft is not a bench of anonymous juniors behind a partner. That is good for judgment. It is a limit if you wanted four squads. A product shop wins that deal. We will say so in the audit.

Training during build is short and specific. The operator runs the happy path, the failure path, and the approval path. They edit a prompt in the place you will keep prompts. They know who to call if a vendor API breaks. Corporate AI training as a separate service is for a wider room. The build includes the operator’s room.

Under 30 people, the founder should see the approval step. Past 250, the department sponsor should see it. Money and customers stay human. That demo is part of done.

• Running workflow in accounts you control
• Source you can build without us on the call
• Prompts and eval notes you can edit
• Operator runbook for Tuesday
• Revocable keys
• Logs you can retain
• Named human approvals

A strategy firm’s leftover is a pack. A boutique’s leftover is a scorecard. A seat vendor’s leftover is a tenant. A shop’s leftover is a repo if the contract says so. Our leftover is the list above. If you only wanted the pack, do not buy a build.

We stay for improve-and-support when you ask. Continuous improvement is cheaper when the first system is simple enough to measure. We will not invent a forever retainer on this page. If you want a monthly senior seat, read rent-an-AI-engineer and still start with a scoped outcome.

We will not build a fully automatic send to customers or a fully automatic payment without an approval step. We will not build against a fantasy API. We will not replace counsel with Protectron-shaped paperwork and call it legal review. We will not take a greenfield consumer app that needs a twenty-person shop. We will not take a company-wide rollout with no first department.

We will not build if the audit (ours or yours) shows the hours live in drafting. Buy Copilot or ChatGPT. We will not build if the only success metric is “we launched AI.” Name a process metric you already believe: time to intake, cost per acquired car, tokens per feature, hours per brief, days from notice to complete file.

We will not pretend a workshop is a build. If you need a playbook, buy training. If you need both, sequence them. The compare page on implementation versus workshops exists so procurement can keep the line items honest.

Inventory acquisition. Nexuvo dealers pull private-party cars with outreach that runs while the desk is closed. 3x more inventory. 80 percent less manual outreach. A build in this class needs a buy box, marketplace access, and a human on the offer.

Operations in a regulated or high-care setting. Omnira Dental is live. Eviction Wizard takes a notice through a writ in one dashboard. Builds in this class need audit trails and a person on filing or clinical exception.

Spend and compliance. TokenTra makes the bill readable by feature. Protectron produces EU AI Act artifacts. Builds in this class need honest invoices and an owner who will act on a flag.

Documents and media. NanoBrief asks industry questions and writes a brief in under two minutes instead of three or more hours. Octana turns a message into a video. Clear Audio AI cleans a recording. MixGuidera suggests EQ and compression. Builds in this class need samples of good output, not a slogan about creativity.

Guidera and Nerlude show two other edges: a companion product a person carries, and a control center a founder checks so renewals and keys do not surprise them. Your SOW might look like neither. The discipline is the same. Name the Tuesday user.

If two rows both look true, pick the leftover you need on the last Friday of the quarter. If that leftover is a pack, do not buy us. If it is a playbook, buy training. If it is a specified app, buy a shop. If it is a live process in your CRM, stay on this page.

We write eval notes for the failure modes we already saw on the floor. Wrong seller. Wrong tooth. Wrong clause. Wrong token route. The operator keeps those notes. When a model version changes, you have something to rerun. This is unglamorous and it is why Own works.

Measures should already exist in the business. Time to complete intake. Cars acquired per week. Tokens per feature. Hours per brief. Days in an eviction file. We will not invent a vanity dashboard that only we understand. If you cannot name a measure, we are still in audit.

Protectron-shaped work needs a compliance owner. TokenTra-shaped work needs a finance or platform owner. Dental and eviction work need the licensed human. If that owner is “the innovation committee,” stop. Committees do not approve a filing at 4pm.

Nidrosoft will not pretend to be a factory. One hundred twenty-five products and sixteen public systems describe a long practice, not parallel armies. If procurement needs twenty CVs, we will fail the sheet. Hire the shop. Keep a studio for the process that still lacks a spec.

Transformation is how more than one team gets a build, in series. Rent-an-engineer is how you keep a senior seat after the first outcome. Neither is a hidden price list. Both start with a map.

Formerly Anthropic, Microsoft, Intuit, and Gap explains the taste for production constraints. It does not mean those companies are the customer of your SOW. Nidrosoft LLC is. Cyriac Zeh signs the work.

Book an audit or send cyriac@nidrosoft.com a paragraph that names the tool, the weekly pain, and the owner. If a strategy firm already wrote a use-case list, send the page numbers. We will mark which items are buildable and which items are still narrative.

If you are in San Diego or La Mesa, we can sit on site for the first embed days. If you are not, we still ship into your systems. We will not invent a staffed office to win a regional RFP.

Done looks like an operator running the path without us, a measure you already believed, and files you can open after we leave. If a vendor offers done as a launch party, keep shopping.

Bring two examples of the process going well and two examples of it failing. Those four records are a better brief than a vision deck. We will use them to write the first eval notes. If you cannot find a failure example, the process may not repeat enough to justify a build. That discovery belongs in the audit, before anyone writes a connector.

If legal already drafted an AI policy, send it. We will mark which clauses affect the write path and which clauses are board language that should stay with whoever wrote them. A build that ignores an existing policy will die in review. A build that treats every clause as a blocker will never ship. The audit is where we sort those clauses.

Octana, Guidera, Clear Audio AI, and MixGuidera are useful only if your process is media or travel. Do not ask us to bolt a video generator onto a dealer desk because it looks modern. Ask us to watch the desk. The public archive is a reference for texture, not a catalog you must buy from.