AI Agent Development: Assistants that do real work, with people in control

Executive summary

Gartner predicts that agentic AI will autonomously resolve 80% of common customer service issues by 2029, contributing to a 30% reduction in operational costs.¹
Gartner also predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing rising costs and unclear business value.²
Gartner estimates that only about 130 of the thousands of vendors claiming agentic AI offer genuine agentic capability, a practice it calls “agent washing”.²
The agents that last have one clear job, access to the right data, defined limits, and a hand-over to a person.
We design agents for enquiries, internal knowledge, reporting and customer support, and measure them like any other team member.

01Market context

Expectations for AI agents are high. In March 2025 Gartner predicted that by 2029 agentic AI will autonomously resolve 80% of common customer service issues without human intervention, leading to a 30% reduction in operational costs.¹ By 2028, Gartner expects at least 15% of day-to-day work decisions to be made autonomously through agentic AI, up from none in 2024.²

Exhibit 1 High expectations, high attrition

Common customer service issues Gartner expects agentic AI to resolve autonomously by 202980%

Agentic AI projects Gartner expects to be cancelled by end of 2027 (more than)40%

Day-to-day work decisions Gartner expects to be made autonomously by 2028 (at least)15%

Source: Gartner, March and June 2025.

The same firm is cautious about today’s projects. In June 2025 it predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027, and estimated that only about 130 of the thousands of vendors marketing “agents” offer the real thing.²

02The challenge

Agents with no clear job. A general assistant is launched and asked to do everything, and does nothing well.
No access to real data. The agent can talk but cannot check availability, an order or a policy.
No limits. Nobody has defined what the agent must never say or promise.
No hand-over. Customers get stuck in a loop instead of reaching a person.
Vendor claims. Products sold as agents are often scripted chatbots.

03Our point of view

A good agent is closer to a well-trained new hire than to software. It has one job description, access to the information it needs, a short list of things it may never do, and a manager to escalate to. We design agents that way: narrow, connected and supervised.

For customer-facing agents, the goal is rarely to replace a person. It is to answer the routine questions instantly, collect the details, and pass a well-prepared case to your team.

04How we build agents

Define the job. One use case, the questions it must handle, and what “resolved” means.
Connect the knowledge. Approved content, policies and system data the agent may use, and nothing else.
Set the limits. Topics it must hand over, words it must not use, promises it must not make.
Test with real questions. Past enquiries are replayed before launch, and failures are fixed.
Launch with oversight. Conversations are reviewed and the agent is improved each week.

Exhibit 2 What every agent needs

A job
One clear use case

Knowledge
Approved data only

Limits
What it must not do

A hand-over
A person to escalate to

Our agent design principles.

05What you receive

A purpose-built assistant for enquiries, internal knowledge, reporting or customer support.
A knowledge base of approved content the agent answers from.
Guardrails and hand-over rules, documented and tested.
A review dashboard showing volumes, resolution and escalations.

06Measures of success

Share of conversations resolved without a person
Accuracy on a sample of reviewed answers
Escalation rate and quality of hand-over
Customer satisfaction after the conversation
Leads or bookings created by the agent

07Risks and governance

Accuracy. Agents answer only from approved content and say when they do not know.
Disclosure. Users are told they are speaking with AI.
Data protection. Personal data is collected only where needed and handled in line with the UAE PDPL.³
Commitments. Prices, refunds and promises stay with people.

This paper is general guidance and not legal advice.

08Illustrative scenario

ILLUSTRATIVE — NOT A CLIENT CASE STUDY

A beauty and wellness lounge receives the same questions every day on WhatsApp and Instagram: prices, availability, location and what a treatment involves.

A booking assistant answers from the lounge’s approved menu and policies, checks open slots, and offers a booking. Questions about allergies or medical conditions go straight to a therapist. The team reviews a sample of conversations each week and updates the knowledge base.

09Frequently asked questions

Is an agent the same as a chatbot?: Not quite. A scripted chatbot follows fixed paths. An agent can understand open questions, use your data and take defined actions, within limits.
Will it make things up?: It is restricted to approved content and is tested and reviewed for accuracy. When unsure, it hands over.
Which channels can it work on?: Typically your website, WhatsApp and internal tools. The channel follows where your customers or staff already are.

AI Agent Development