Ivan Pylypchuk (@IvanSoftblues) header
Ivan Pylypchuk (@IvanSoftblues) profile picture

AI systems architect, practitioner & entrepreneur. CEO, Softblues. I design AI systems that run a business. Claude, ChatGPT, Copilot and custom agents.

London, United Kingdom
Joined October 2019
Photo by Ivan Pylypchuk (@IvanSoftblues)Photo by Ivan Pylypchuk (@IvanSoftblues)Photo by Ivan Pylypchuk (@IvanSoftblues)Photo by Ivan Pylypchuk (@IvanSoftblues)Photo by Ivan Pylypchuk (@IvanSoftblues)Photo by Ivan Pylypchuk (@IvanSoftblues)Photo by Ivan Pylypchuk (@IvanSoftblues)Photo by Ivan Pylypchuk (@IvanSoftblues)Photo by Ivan Pylypchuk (@IvanSoftblues)Photo by Ivan Pylypchuk (@IvanSoftblues)
A tidy spreadsheet can hide a bad extraction. For a document-to-table trial, keep the source page and exact text beside each value. Missing renewal date? Leave it blank and flag “not found”. Have a person check the row before it reaches a live system. Measure extraction + review + corrections, not copying time alone. The image shows the review flow. Process calculator in the reply.
1
16
Before rolling out a policy assistant, test a question it cannot answer. With a synthetic expenses policy, check: • A covered question: answer + exact source paragraph + version. • A missing rule: “Not covered by this policy.” • An unauthorised account: no document contents retrieved. Open the citation and check it. Permissions belong in the system, not the prompt. Keep these tests for the next policy revision.
1
8
I changed one input for a 60-person firm: planned hires. Same AI capacity: 5.5 people's time by month 12. 0 hires: £0 added to profit; no payback within 24 months. 3 hires that can wait: £135,000 a year from month 12; payback month 14. Illustrative starting assumptions for a typical UK firm at typical prices, not a client result. For a broker, check which team gets the time before treating a hire as avoidable. Calculator in the reply.
1
7
Give AI your approved letter template, a brief and a style example. Tell it which wording must stay fixed and which facts it must never invent. Synthetic example: "Please send the signed form by [MISSING: deadline] to [MISSING: return address]." Ask for the draft plus a gap list. A person checks the facts, resolves gaps and sends. Once tested, save the method as a reusable Skill; keep client details separate. Our Claude playbook covers this as Agent 3, P1. Link in the reply.
1
13
Choose the Claude guide for the free 21-slide deck, including the house-template example. Use invented details for your first test: softblues.io/ai-adoption/gui…
11
Suppose you check 200 document packs a month. AI cuts human work from 20 to 8 minutes per pack, including review. (20 - 8) × 200 ÷ 60 = 40 hours freed. An illustrative example, not a client result. Measure your own process first. Calculator below.
1
11
Your reviewer should not have to hunt for the evidence behind an AI finding. Softblues' published financial-advice case proposes around 40 checks per file, each linked to a source page. It is pre-discovery, with no measured results. A person keeps the final decision. Test a known problematic file before buying a build. Count the issues missed and the time spent correcting the output, not just how quickly a summary appears.
1
21
Your next AI investment needs a first business case. Softblues' Business AI Adoption calculator asks five questions, then shows typical tasks for your industry, a staged adoption scenario and potential payback. The numbers are illustrative UK assumptions at typical prices, not client results. Change the inputs to match your firm, try the cautious case and take the board page to your leadership meeting. Free, no sign-up.
1
8
Your research task can use millions of tokens behind a short answer. Our Astra research trace recorded 10.14M tokens. Its Standard API token equivalent was about $12.77, not a subscription bill. OpenRouter reports that nearly 90% of agentic tokens were cached in September. Price the whole run: cached input, new input and output. Then add the time spent checking what passed.
1
18
Your next AI connection should come with a demonstration of what it cannot change. I would test it on one ordinary job: prepare a client review pack. Find the engagement letter, read the correspondence, check the outstanding task and draft the note. Then stop before sending it or changing the client record. Microsoft's "Configure data policies for agents", checked on 12 October, says that blocking a Power Platform connector also blocks tools in the MCP server using it. The actual connector matters. Blocking an unrelated social connection proves nothing about the new one. Claude and ChatGPT also separate access from action. Do not assume an administrator's Microsoft consent settles which operations the assistant can use or when a person confirms them. My first step would cost no new software budget: ask IT to demonstrate a permitted read and a refused change under the intended staff account. Then test the workflow with existing licences. Add another assistant or commission a connection only when the result justifies it. The full comparison includes platform limits and a 100-person cost table, with UK prices and separately labelled dollar subscriptions: softblues.io/blog/mcp-server…
1
44
A useful connector test has an expected refusal. Use a record the test account must not change, attempt the operation in a safe test environment and keep the result. A successful demo alone tells you little about the boundary.
1
21
Compare assistants on the accepted client pack: missing sources, corrections and review effort. Keep sending and changes to the client record with the responsible person until the approval path has been tested.
3