Consulting wants to become software again

The Anthropic and Tata Consultancy Services announcement is a better autonomous-business signal than another demo of an agent booking a meeting. TCS says it will put Claude in front of 50,000 employees across 56 countries, then use that internal learning to build Claude-powered products for clients in financial services, healthcare, public sector, aviation, telecom, life sciences, and other regulated industries.

The interesting word is not products. It is products after customer zero. TCS is using itself as the rehearsal space, then packaging what works. That is the oldest services-to-software story in the book, just with agents as the labor layer instead of junior analysts and custom workflow glue.

The examples are not vague either. Anthropic says Diligenta, TCS's UK life and pensions business, will use Claude to improve customer experience for more than 22 million policyholders. Banking and financial services teams will use Claude Code for software engineering and IT operations. TCS engineering teams will add reusable skills and plugins to Claude Code, starting around claims adjudication and lending advisory. TCS iON, which runs more than 75 million assessments each year across 1,500 Indian cities, will deliver training and certification.

The new services product is an operating loop

A normal consulting firm sells hours, decks, and transformation incense. An agent-native services firm sells a loop: intake, context, execution, evaluation, exception handling, reporting, and improvement. The staff still matters. They are no longer only doing the work. They are designing the rails the work runs on.

That is why regulated industries show up so early. Insurance claims, lending advice, healthcare operations, public-sector systems, and compliance-heavy finance are packed with repeatable work that cannot simply be handed to a cheerful black box. The buyer needs provenance, auditability, policy, escalation, and someone to blame if the machinery gets ambitious. A services firm already has the trust wrapper. The opportunity is to turn that wrapper into repeatable agent infrastructure.

For an autonomous business builder, the lesson is sharp: do not sell 'AI automation.' Sell the named workflow. Claims adjudication. Lending packet review. Vulnerability triage. Renewal operations. Permit intake. Lead qualification. Each named workflow can become a product if the context, tools, review gates, and output standard are stable enough.

Voice is becoming the front desk

OpenAI's GPT-Live announcement matters for this same reason. The product details are about natural conversation: a full-duplex architecture that can listen and speak at the same time, respond without rigid turn-taking, and delegate deeper work to a frontier model in the background. OpenAI says GPT-Live will power ChatGPT Voice, roll out globally, and come to the API later.

The business implication is not that every company needs a charming voice bot. Please, no. The implication is that the front door to work is getting more ambient. A customer, employee, or field operator can speak to a system that stays present while another model searches, reasons, or runs a task behind the curtain.

That starts to make old service desks look strange. Why should a regulated-industry workflow begin with a form if the system can hold a conversation, collect missing fields, explain what it is doing, and route the case to a background agent? The form does not vanish. It becomes the artifact produced by the conversation.

The measurement layer is not optional

OpenAI's July 8 coding-evaluations post is the cold shower this market needs. The company audited SWE-Bench Pro and estimated that roughly 30% of the tasks were broken. It found overly strict tests, underspecified prompts, low-coverage tests, and misleading prompts. On the 731-task public split, frontier model pass rates had jumped from 23.3% to 80.3% in eight months, but the measurement instrument itself had serious defects.

That is not only a benchmark story. It is an operating story. If your autonomous business cannot tell the difference between a completed task and a task that merely passed a brittle check, you do not have leverage. You have theater with invoices.

The services firms that win here will build evaluation into the product, not into the quarterly review deck. Every claims packet, remediation patch, support resolution, lead handoff, and compliance memo needs a quality trail. What was the input? What context was used? What did the agent decide? What did a human approve? What failed? What changed next time?

Open source pulls the model downstream

OpenClaw's v2026.7.1-beta.3 tag is a small data point, but the project is worth watching because it compresses this enterprise pattern into the local-first, multi-channel, open-source world. The lesson is not that every business should run a fresh beta release in production. The lesson is that persistent agents are moving from vendor dashboards into messier real-world channels.

Once the agent can receive work through chat, terminal, schedules, and local tools, the product boundary changes. The company no longer sells a screen. It sells a way for work to enter, be handled, be checked, and be remembered.

That is where small autonomous businesses get interesting. A five-person shop can imitate the shape of TCS at one narrow altitude: pick one regulated or semi-regulated workflow, build a channel, wrap it in policy, instrument it hard, and sell the loop.

The operator takeaway

If you want to build from this, start with a services business that already has ugly repetitive work and expensive trust requirements. Do not begin with a model. Begin with the before-and-after ledger: what arrives, who owns it, what counts as done, what must be checked, what evidence proves the check happened, and what the customer sees.

Then make the workflow smaller until it can be packaged. A workflow that needs heroic judgment on every case is still consulting. A workflow that follows a narrow playbook, escalates exceptions, and improves its own playbook can become an autonomous business.

The services firm is not dead. It is turning into a factory for agent-operated products. The factory floor is made of prompts, permissions, evals, integrations, logs, and humans with enough taste to stop the line.

The next consultancy pitch should come with a dashboard, a test suite, and a pause button.

That would be less glamorous than a vision deck. It would also be much closer to a business.

sources
  1. TCS and Anthropic partner to bring Claude to regulated industriesAnthropic, accessed July 9, 2026
  2. Introducing GPT-LiveOpenAI, accessed July 9, 2026
  3. Separating signal from noise in coding evaluationsOpenAI, accessed July 9, 2026
  4. OpenClaw v2026.7.1-beta.3 releaseGitHub, accessed July 9, 2026