Can it work with our existing systems?

Usually, yes. We connect through documented APIs or MCP servers (HubSpot, Zoho, Shopify and Google Workspace, among many others, publish them) and build thin, tested connectors where none exist.

Is our data used to train the models?

Under Anthropic's and Google Cloud's commercial API terms, your inputs and outputs are not used to train their models by default. We confirm retention and data-region settings for your project in writing.

Will people know they are talking to an AI?

Yes. Our agents always say they are AI and offer a route to a person. Anthropic's usage policy requires this for consumer-facing assistants, and we would insist on it anyway.

AI and agents

AI agents and automation

Agents that answer, act and hand off, on Claude and Gemini, inside the tools you already run.

Book a call

What we deliver

We design and ship production AI agents on Claude and Gemini: customer-facing assistants, back-office agents that work across your CRM, ERP and inbox, and automations with guardrails and human approval built in.

What we build

Customer and sales assistants on your website and WhatsApp that answer from your own content, look up orders and bookings, qualify enquiries and book calls.

Back-office agents that read email, PDFs and spreadsheets, draft replies and documents, and update your systems, with a person approving anything that matters.

Internal copilots that let your team ask questions of policies, contracts and data in plain language.

How we make agents dependable

Every agent gets a small set of well-defined tools with strict schemas, so it can only do what you have allowed. Text from the outside world (emails, web pages, form fields) is treated as data, never as instructions.

We write evaluation suites before launch, from realistic conversations and edge cases scored automatically, and keep running them as prompts, models and data change.

Handing over to a person is a designed path, not an afterthought, and every action is logged so you can see what the agent did and why. Pip, Softpiper's AI on this site, runs on Claude with exactly these patterns.

Claude, Gemini or both

We choose models per task: Claude for long, tool-heavy reasoning and careful writing; Gemini for real-time voice, video understanding and generative media; small, fast models for routing and classification.

On Google Cloud both families sit side by side in Gemini Enterprise Agent Platform (formerly Vertex AI), so you can keep one cloud and one bill. Claude is also available through Anthropic's API, AWS and Microsoft Azure.

Typical deliverables

Use-case discovery and agent design
Tool and MCP integrations with your CRM, ERP, commerce and calendar
Guardrails, permissions and human approval steps
Evaluation suite and quality reporting
Rollout on web, WhatsApp, Slack or email
Monitoring, cost tracking and runbooks
Team training and handover

Tools we typically use

Claude (Opus, Sonnet, Haiku)
Gemini, including Gemini Live for voice
Model Context Protocol (MCP)
Claude Agent SDK
Google Agent Development Kit (ADK)
Gemini Enterprise Agent Platform
WhatsApp Business Platform
Postgres and Supabase
Next.js

If your team already runs a different stack, we work within it. These are defaults, not requirements.

Claude and Gemini

Where Claude and Gemini fit

We build on both model families and choose per task, on evaluation results, cost and latency, inside the cloud you already use.

Claude

by Anthropic

Long, tool-heavy work: reading email and documents, calling your CRM, ERP or commerce systems through MCP tools with strict schemas, and drafting replies for a person to approve.

Gemini

by Google

Real-time voice agents with Gemini Live, multilingual calls and transcription, and agents that need to understand images or video.

Together. A small, fast model routes each request, Claude reasons and acts, Gemini handles voice and media, and a person approves anything consequential.

Phases and team

How an engagement typically runs

6–12 weeks to a production pilot, delivered by a small senior pod, with progress you can see as it is built. Pip tailors the working model to your team.

Discover
Map the conversations or workflows worth automating, the systems involved and where a person must approve.
Prototype
A working agent on your real content and tools, with an evaluation suite built from realistic cases.
Pilot
A production pilot with guardrails, human approval steps, logging and cost tracking.
Scale and manage
Wider rollout, evals on every change and ongoing management as models and your data evolve.

Typical team

Solution architect, AI engineer, full-stack engineer and QA/evals engineer

Pricing

Priced to your scope and market.

Every project is priced to its scope and market. Pip gives you a tailored range in minutes, and the final number comes in a written proposal.

Hosting, subscriptions, licences and AI usage are billed to your own accounts at cost, estimated up front.

How pricing works

Related work

Where we've done this

All work

Our productLive

LuxeLens AI Studio

A self-serve studio, built and operated by Softpiper, that turns one product photo into styled product shots, on-model images and short video clips using Google's Gemini image models and Veo 3.1.

Product, Model and Jewellery studio modes
AI Director agent: brief to shot plan (camera, lens, lighting, set, grade)
Product Identity Lock grounded with Google Search

Read the case study Open the product

FAQ

AI agents: common questions

Ask Pip if yours isn't here.

We design against it. Answers are grounded in your own content and systems, tools return structured data, and the agent is instructed to say when it doesn't know. Evaluation suites catch regressions before each release. No model is perfect, so consequential actions always pass through a human approval step.

Related services

All services

Pip · Softpiper's AI

Ask Pip

Get a tailored range in a couple of minutes

Pip is Softpiper's AI. Describe the project and it will ask a few questions, give you a tailored price range for your scope and market, and book a call with the team.